Smooth Reparameterizations of Functions on Simplicial Product Spaces: Applications to Probabilistic Tensor Decomposition and Functional Data Registration

Anuj Srivastava; Arafat Rahman; P.-A. Absil; Shashwat Kumar

REVIEW 2 major objections 5 minor 32 references

Smooth Reparameterizations of Functions on Simplicial Product Spaces: Applications to Probabilistic Tensor Decomposition and Functional Data Registration

T0 review · 2 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Simplex-constrained optimization can be replaced by a smooth manifold problem using any elementwise strictly convex map ψ with ψ(0)=0 and ψ''>0, and this replacement preserves weak second-order KKT points.

desk verdict Solid incremental generalization of the Hadamard KKT result, with a real flaw in Remark 4.1 and a weak practical optimality certificate. read the letter →

arxiv 2608.02576 v1 pith:SMUTKRF6 submitted 2026-08-03 cs.LG

Shashwat Kumar , Arafat Rahman , Anuj Srivastava , P.-A. Absil This is my paper · ORCID

classification cs.LG MSC 65K0553B2170G45

keywords RiemannianOptimizationTensorDecompositionFunctionalDataRegistrationHadamardParametrizationproductsimplexweaksecond-orderKKTSRVFreparameterization

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

Many practical optimization problems pin variables to products of probability simplices, such as low-rank discrete probability tensors and the warping functions used to align functional curves. The paper shows that these non-smooth constraints can be replaced by a smooth manifold using any elementwise map ψ(θ) with ψ(0)=0 and ψ''>0, and that this replacement preserves the correct optimality notion: weak second-order KKT points of the simplex problem correspond exactly to second-order KKT points of the manifold problem. The proof runs through a Hessian decomposition that separates a positive diagonal boundary term from the pulled-back Hessian of the objective. On the applied side, the paper introduces Riemannian gradient descent, including a spectral-step variant, and reports faster convergence than projected gradient descent on tensor factorization and smoother, more accurate warping functions in square-root-velocity functional data registration.

What carries the argument

The engine is the reparameterization class Ψ = {ψ : R → R_{≥0} | ψ(0)=0, ψ''>0}, applied coordinatewise to each simplex block; the coordinatewise square map ψ(θ)=θ² is the special case, but the class also includes even polynomials with positive coefficients, cosh θ−1, and Legendre transforms of such maps. The key identity is a Hessian formula showing that the manifold Hessian splits into a diagonal term carrying ψ''(θ_l) times the KKT multiplier β_l, plus the objective Hessian pulled back through the Jacobian of φ. Because ψ''>0 and β_l≥0 at simplex KKT points, this diagonal term cannot make the quadratic form negative; the reverse direction uses tangent vectors supported on zero coordinates

What would settle it

On the 2-simplex in R^3 with a quadratic objective and ψ(θ)=θ², pick a boundary point with exactly one active zero coordinate. Verify whether every preimage θ* satisfies the manifold second-order condition 1^T(ψ'(θ*)⊙Δθ)=0 ⇒ Δθ^T∇²L2Δθ≥0, while the simplex weak-KKT quadratic form is nonnegative on its tangent set. A single point where the two signs disagree disproves the claimed equivalence.

Watch

Extended reading notes

Core claim

The central claim is an optimality correspondence between the constrained problem on a product simplex and the reparameterized problem on a smooth manifold. For every ψ in the allowed class, the image of every first-order KKT point of the simplex problem is a first-order KKT point of the manifold problem, and the image of every weak second-order KKT point is a second-order KKT point. Conversely, every second-order KKT point of the manifold problem maps to a weak second-order KKT point of the simplex problem. The paper stresses that only the weak form is preserved: the strong second-order condition can fail on the simplex even when the manifold point is second-order critical, and every vertex

Load-bearing premise

The equivalence leans on reparameterizations with ψ(0)=0 and ψ''>0, which force ψ and ψ' to vanish only at zero so that zero simplex coordinates correspond to zero manifold coordinates; the registration experiments additionally assume the discrete warping-rate problem faithfully represents the continuous problem after relaxing ˙γ>0 to ˙γ≥0.

Editorial extensions

If this is right

Any Riemannian optimization algorithm that converges to a second-order KKT point on the smooth manifold returns a point that meets the weak second-order necessary conditions of the original product-simplex problem.
The KKT-preserving behavior is intrinsic to the whole class ψ(0)=0, ψ''>0, not a special feature of the coordinatewise square map, giving practitioners freedom to choose reparameterizations that improve conditioning.
The correspondence extends to feasible sets that are products of several simplices plus Euclidean blocks, covering simplex-constrained tensor factorization and SRVF-based functional data registration.
In the numerical experiments, Riemannian gradient descent with a spectral step reaches lower objective values in less time than projected gradient descent, and the resulting warping functions are smoother and closer to the true shapes.
The strong second-order KKT condition is not preserved, and every simplex vertex is a weak second-order KKT point, so manifold algorithms can terminate at a vertex without the weak condition certifying genuine local optimality.

Reading between the lines

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

A natural stress test is to pick ψ(θ)=θ⁴+θ² or ψ(θ)=cosh θ−1 in the same two applications and check whether the practical speed advantage persists; the theory guarantees the KKT correspondence, but not the conditioning advantage, so this separates the theorem from the algorithm.
Because the proof treats f as any twice-differentiable function, the result should transfer to other product-simplex problems such as hyperspectral unmixing or nonnegative matrix factorization; the same Riemannian recipe is a plausible drop-in replacement for projected methods there.
The discretized registration problem relaxes strict warp monotonicity ˙γ>0 to ˙γ≥0; a follow-up consistency check would be whether the recovered discrete warping functions, when interpolated, remain strictly increasing, or whether a strict-positivity regularizer is needed in practice.

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

Dear colleague,

The short version: the paper does what it says — extends the Hadamard-parameterization KKT result of Li et al. to product simplices with Euclidean blocks and to any elementwise strictly convex ψ with ψ(0)=0 and ψ''>0. The main theorems (4.1–4.4) are sound as far as I can tell. I stress-tested the multiplier arguments, the tangent-space logic, and the Hessian decomposition (10), and found no fatal gap. The weak/strong KKT distinction and the Legendre-transform family are genuinely new, though the proof strategy closely follows Li et al. That is incremental but legitimate work.

The softer spots, in order of importance. First, the advertised 'weak second-order KKT' certificate is almost vacuous at vertices. At a vertex, the tangent constraints (6c) force Δv=0, so the curvature inequality holds automatically whenever first-order KKT holds. Thus a local maximum at a vertex passes the certificate. This is not a contradiction of the theorem, but it means the practical optimality guarantee is much weaker than a local-minimality certificate. Relatedly, Remark 4.1 overstates: it claims every vertex is weak second-order KKT regardless of objective. That is false because first-order KKT must hold; e.g., f(v)=v1 on Δ² at (1,0,0) violates β≥0. The correct statement is that any vertex that is first-order KKT is trivially weak second-order KKT. This should be fixed.

Second, the numerical comparisons are presented without error bars or statistical tests. The 'outperforms PGD' claim rests on a few hand-tuned step sizes and no variance; for a paper whose theory is the main contribution, this is not disqualifying but the empirical claims should be toned down. Third, the discretized SRVF registration relaxes γ̇>0 to γ̇≥0, so the warping functions can have flat spots; the authors acknowledge this but it means the product-simplex model is an approximation to the true diffeomorphic registration problem. Minor issues: Theorem 4.3's 'absorbing g_l in β_l' step is informal (though it can be made rigorous), and Eq. (29) has a subscript typo.

Who should read this: people working on simplex-constrained optimization who want to know whether they can replace the simplex with a smooth manifold and keep KKT guarantees. For that audience, it is useful and deserves serious refereeing. My recommendation: send it to a good optimization journal, require a corrected Remark 4.1, a more careful statement of what the weak certificate does and does not imply, and better uncertainty treatment in the experiments. I'd take it in a reading group, and I'd cite it if I were writing about reparameterization tricks.

Referee Report

2 major / 5 minor

Summary. The paper proposes a general smooth reparameterization of optimization problems over product spaces of simplices, using elementwise maps ψ: R→R≥0 with ψ(0)=0 and ψ''>0, of which the Hadamard map is a special case. The central theoretical result is a KKT correspondence: weak second-order KKT points of the original simplex-constrained problem (P_simplex) map to second-order KKT points of the reparameterized manifold problem (P_×smooth) (Theorem 4.2), and conversely every second-order KKT point of the manifold problem maps to a weak second-order KKT point of the simplex problem (Theorems 4.3–4.4). The authors apply this framework to simplex-constrained canonical polyadic tensor decomposition and SRVF-based functional data registration, and present numerical experiments with Riemannian gradient descent (RGD) variants outperforming projected gradient descent (PGD) in both settings.

Significance. If the theoretical correspondence is correct, it extends the known Hadamard-parameterization results to a broader class of smooth reparameterizations and to product simplices with Euclidean components, which is a meaningful generalization for tensor factorization and functional data analysis. The proofs are self-contained, the distinction between weak and strong second-order conditions is clearly articulated, and the accompanying code is a useful asset. The numerical results are suggestive but not statistically validated, and the practical strength of the weak second-order certificate is limited, especially at boundary points. Nevertheless, the core KKT mapping appears sound and the paper makes a credible contribution to constrained optimization on simplex-structured feasible sets.

major comments (2)

[§4, Remark 4.1] The statement that 'regardless of the objective function, every vertex of the simplex is a weak second-order KKT point' is false as written. A weak second-order KKT point must first satisfy the first-order KKT conditions (4). For f(v)=v1 on Δ^2 at v=(1,0,0), the condition (4a) with β⊙v=0 forces β2=-1, violating β≥0. Thus the vertex is not even first-order KKT. The correct statement is that every vertex that satisfies first-order KKT automatically satisfies the weak second-order condition because (6c) forces Δv=0. This should be corrected, and the dependence of the claim on first-order feasibility should be acknowledged.
[§4, Eq. (6) and Thms 4.2–4.4] The advertised guarantee is nearly vacuous at boundary points. Condition (6c) forces Δv_i=0 for every coordinate where v_i=0, regardless of the multiplier β_i. At a vertex, this leaves Δv=0, so Δv^T ∇^2 L1 Δv = 0 is an identity. Consequently, the 'weak second-order KKT' certification provides no local-minimality information at vertices; it only says that no negative curvature is detected along the (empty) admissible tangent cone. The paper should either prove a nontrivial result involving the strong second-order condition on a subset of tangent directions, or explicitly discuss this limitation when interpreting Theorems 4.2–4.4. As it stands, the abstract's phrase 'second-order KKT points' may overstate the practical strength of the guarantee for boundary solutions.

minor comments (5)

[§4, Theorem 4.3] The step 'Absorbing g_l in β_l' between Eqs. (21) and (22) is informal. The residual g_l should be defined precisely, e.g., β̂_l = β_l + g_l, and the sign conventions checked before proceeding to the complementarity argument. The proof is recoverable but requires more care.
[§4, Theorem 4.4, Eq. (29)] There is a subscript typo: the expression '∆θ^T ∇^2_v L2 ∆θ' should be '∆θ^T ∇^2_θ L2 ∆θ', since the Hessian is taken with respect to θ, not v.
[§6, Figs. 2–3] The numerical comparisons report objective values without error bars, standard deviations, or statistical tests. For the tensor experiments, the number of trials is not stated; the registration section mentions 50 trials, but the plots would benefit from showing variability across seeds. Adding such information would strengthen the empirical claims.
[§5.2, Eq. (50)] The relaxation from γ̇>0 to γ̇≥0 is essential for the problem to fit the product-simplex framework, but its effect on the registration solution is not discussed. The KKT theory applies to the relaxed problem, and the paper should clarify that the original diffeomorphism constraint is only approximately represented.
[Algorithms 1–2] The pseudocode and the algorithms work with the Hadamard parameterization specifically (line 'θ⊙θ'), whereas the theoretical development allows any ψ satisfying (1). If the numerical experiments use only ψ(θ)=θ², this should be stated explicitly; otherwise the algorithms should be written for general ψ.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the KKT correspondence is proved from explicit assumptions; the main caveat is a non-circular overstatement in Remark 4.1 about vertices.

full rationale

The main derivation chain (Theorems 4.1–4.4) is self-contained. The admissible reparameterizations are fixed by assumption (1), the KKT systems are laid out in (4)–(7), the Hessian decomposition is given in (10), and each implication is proved by explicit construction: Theorem 4.2 maps admissible Δv to smooth tangent vectors Δθ and uses β≥0 and ψ''>0; Theorem 4.3 uses special coordinate tangent vectors at zero coordinates to force β̂≥0; Theorem 4.4 builds Δθ from Δv and evaluates the Hessian. No fitted parameter is renamed as a prediction, and the numerical sections are benchmarks against PGD rather than tests of a fitted quantity. The self-citation [19] (and background refs [27,28]) supplies only the SRVF application setup; it is not load-bearing for the KKT equivalence. One caveat belongs to correctness, not circularity: Remark 4.1 claims that every vertex is a weak second-order KKT point 'regardless of the objective function,' but a vertex must first satisfy first-order KKT; e.g., f(v)=v1 on Δ² at (1,0,0) is not first-order KKT. This makes the boundary curvature certificate vacuous though valid, but it is not a case of the derivation reducing to its inputs. Hence the score reflects only a minor, non-load-bearing self-citation, not actual circularity.

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

No parameters are fitted in the theory. The only assumptions are the smoothness of f and the class of strictly convex ψ defined in (1). The numerical experiments involve hand-tuned hyperparameters, but these do not affect the central KKT claim.

assumptions (3)

standard math f is twice continuously differentiable on the feasible set
Stated in the Introduction ('a twice differentiable function f') and required for second-order KKT conditions and Hessian manipulations.
domain assumption ψ: R→R≥0 satisfies ψ(0)=0 and ψ''>0
Equation (1) defines the allowed reparameterization class. Strict convexity and zero-at-zero are essential for the boundary degeneracy θ=0↔v=0 on which the KKT proofs rely.
domain assumption The feasible set is a product of simplices and Euclidean spaces
This is the problem class under study, stated in (P_simplex). The theorems extend the single-simplex setting to this product structure.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Smooth Reparameterizations of Functions on Simplicial Product Spaces: Applications to Probabilistic Tensor Decomposition and Functional Data Registration." pith.science (2026). https://pith.science/paper/SMUTKRF6

@misc{pith2026260802576,
  author       = {Pith},
  title        = {Pith review of: Smooth Reparameterizations of Functions on Simplicial Product Spaces: Applications to Probabilistic Tensor Decomposition and Functional Data Registration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SMUTKRF6}},
  note         = {Machine review of arXiv:2608.02576}
}

read the original abstract

We consider optimization problems defined on product spaces of simplices. Examples of this class of problems include learning low-rank discrete multivariate probability distributions via simplex constrained tensor decomposition and performing functional data registration under the Square Root Velocity Function (SRVF) representation. In this work, we demonstrate the feasibility of replacing the product simplex with a smooth, elementwise strictly convex reparameterization, resulting in an unconstrained optimization problem on a manifold. We show that performing such a reparameterization results in the second order Karush-Kuhn-Tucker (KKT) points on the smooth manifold being mapped to the weak second order KKT points on the product simplex. This leads to a Riemannian Gradient Descent (RGD) algorithm for solving the reparameterized problem, which outperforms Projected Gradient Descent (PGD), and provides a more faithful representation of the original function shapes while performing curve registration.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 2 canonical work pages

[1]

SIAM Journal on Mathematics of Data Science3(2), 593–623 (2021)

Abdolali, M., Gillis, N.: Simplex-structured matrix factorization: Sparsity-based iden- tifiability and provably correct algorithms. SIAM Journal on Mathematics of Data Science3(2), 593–623 (2021)

2021
[2]

Princeton University Press, Princeton, NJ (2008)

Absil, P.A., Mahony, R., Sepulchre, R.: Optimization Algorithms on Matrix Manifolds. Princeton University Press, Princeton, NJ (2008). URLhttp://press.princeton.edu/ titles/8586.html

2008
[3]

SIAM Journal on Optimization22(1), 135–158 (2012)

Absil, P.A., Malick, J.: Projection-like retractions on matrix manifolds. SIAM Journal on Optimization22(1), 135–158 (2012). DOI 10.1137/100802529. URLhttp://link. aip.org/link/?SJE/22/135/1

work page doi:10.1137/100802529 2012
[4]

IEEE Transactions on Signal Processing69, 6195–6205 (2021)

Amiridi, M., Kargas, N., Sidiropoulos, N.D.: Information-theoretic feature selection via tensor decomposition and submodularity. IEEE Transactions on Signal Processing69, 6195–6205 (2021)

2021
[5]

Society for Industrial and Applied Math- ematics (2019)

Arag´ on, F.J., et al.: Nonlinear Optimization. Society for Industrial and Applied Math- ematics (2019)

2019
[6]

Journal of the Operational Research Society 48(3), 334–334 (1997)

Bertsekas, D.P.: Nonlinear programming. Journal of the Operational Research Society 48(3), 334–334 (1997)

1997
[7]

SIAM Journal on Optimization10(4), 1196–1211 (2000)

Birgin, E.G., Mart ´ ınez, J.M., Raydan, M.: Nonmonotone spectral projected gradient methods on convex sets. SIAM Journal on Optimization10(4), 1196–1211 (2000)

2000
[8]

Springer Science & Business Media (2006)

Bonnans, J.F., Gilbert, J.C., Lemar´ echal, C., Sagastiz´ abal, C.A.: Numerical optimiza- tion: theoretical and practical aspects. Springer Science & Business Media (2006)

2006

Show all 32 references

[9]

Cambridge Uni- versity Press (2023)

Boumal, N.: An introduction to optimization on smooth manifolds. Cambridge Uni- versity Press (2023). DOI 10.1017/9781009166164. URLhttps://www.nicolasboumal. net/book

2023 doi
[10]

In: Proceedings of the 25th international conference on Machine learning, pp

Duchi, J., Shalev-Shwartz, S., Singer, Y., Chandra, T.: Efficient projections onto thel1- ball for learning in high dimensions. In: Proceedings of the 25th international conference on Machine learning, pp. 272–279 (2008)

2008
[11]

arXiv preprint arXiv:2503.24075 (2025)

Esposito, F., Ang, A.: Riemannian multiplicative update for sparse simplex constraint using oblique rotation manifold. arXiv preprint arXiv:2503.24075 (2025)

2025 arXiv
[12]

Physica D: Nonlinear Phenomena53(2-4), 217–232 (1991)

Faybusovich, L.: Hamiltonian structure of dynamical systems which solve linear pro- gramming problems. Physica D: Nonlinear Phenomena53(2-4), 217–232 (1991)

1991
[13]

SIAM (2019)

Gill, P.E., Murray, W., Wright, M.H.: Practical optimization. SIAM (2019)

2019
[14]

Journal of Optimization Theory and Applications 158, 33–64 (2013)

Guo, L., Lin, G.H., Ye, J.J.: Second-order optimality conditions for mathematical pro- grams with equilibrium constraints. Journal of Optimization Theory and Applications 158, 33–64 (2013). DOI 10.1007/s10957-012-0228-x

2013 doi
[15]

IEEE Transactions on Geoscience and Remote Sensing60, 1–13 (2021)

Guo, Z., Min, A., Yang, B., Chen, J., Li, H., Gao, J.: A sparse oblique-manifold nonnega- tive matrix factorization for hyperspectral unmixing. IEEE Transactions on Geoscience and Remote Sensing60, 1–13 (2021)

2021
[16]

IMA Journal of Numerical Analysis38(1), 495–517 (2018) 22 Shashwat Kumar et al

Iannazzo, B., Porcelli, M.: The Riemannian Barzilai–Borwein method with nonmono- tone line search and the matrix geometric mean computation. IMA Journal of Numerical Analysis38(1), 495–517 (2018) 22 Shashwat Kumar et al

2018
[17]

Kolmogorov extension

Kargas, N., Sidiropoulos, N.D., Fu, X.: Tensors, learning, and “Kolmogorov extension” for finite-alphabet random vectors. IEEE Transactions on Signal Processing66(18), 4854–4868 (2018)

2018
[18]

IEEE Transactions on Geoscience and Remote Sensing55(9), 4925–4943 (2017)

Kizel, F., Shoshany, M., Netanyahu, N.S., Even-Tzur, G., Benediktsson, J.A.: A stepwise analytical projected gradient descent search for hyperspectral unmixing and its code vectorization. IEEE Transactions on Geoscience and Remote Sensing55(9), 4925–4943 (2017)

2017
[19]

Plos one20(10), e0332,383 (2025)

Kumar, S., Rahman, A., Gutierrez, R., Livermon, S., McCrady, A.N., Blemker, S., Scharf, R., Srivastava, A., Barnes, L.E.: A shape-based functional index for objective assessment of pediatric motor function. Plos one20(10), e0332,383 (2025)

2025
[20]

arXiv preprint arXiv:2207.03512 (2022)

Levin, E., Kileel, J., Boumal, N.: The effect of smooth parametrizations on nonconvex optimization landscapes. arXiv preprint arXiv:2207.03512 (2022)

2022 arXiv
[21]

arXiv preprint arXiv:2112.05273 (2021)

Li, Q., McKenzie, D., Yin, W.: From the simplex to the sphere: Faster constrained optimization using the Hadamard parametrization. arXiv preprint arXiv:2112.05273 (2021)

2021 arXiv
[22]

IEEE Signal Processing Magazine31(1), 67–81 (2013)

Ma, W.K., Bioucas-Dias, J.M., Chan, T.H., Gillis, N., Gader, P., Plaza, A.J., Ambikap- athi, A., Chi, C.Y.: A signal processing perspective on hyperspectral unmixing: Insights from remote sensing. IEEE Signal Processing Magazine31(1), 67–81 (2013)

2013
[23]

Niculescu, C., Persson, L.E.: Convex functions and their applications, vol. 23. Springer (2006)

2006
[24]

Springer (2006)

Nocedal, J., Wright, S.J.: Numerical Optimization, 2nd edn. Springer (2006)

2006
[25]

IEEE Transactions on signal processing65(13), 3551–3582 (2017)

Sidiropoulos, N.D., De Lathauwer, L., Fu, X., Huang, K., Papalexakis, E.E., Faloutsos, C.: Tensor decomposition for signal processing and machine learning. IEEE Transactions on signal processing65(13), 3551–3582 (2017)

2017
[26]

IEEE Transactions on Signal Processing (2023)

Sidiropoulos, N.D., Karakasis, P.A., Konar, A.: Minimizing low-rank models of high- order tensors: Hardness, span, tight relaxation, and applications. IEEE Transactions on Signal Processing (2023)

2023
[27]

Srivastava, A., Klassen, E.P.: Functional and shape data analysis, vol. 1. Springer (2016)

2016
[28]

Computational Statistics & Data Analysis61, 50–66 (2013)

Tucker, J.D., Wu, W., Srivastava, A.: Generative models for functional data using phase and amplitude separation. Computational Statistics & Data Analysis61, 50–66 (2013)

2013
[29]

Advances in Neural Information Processing Systems32(2019)

Vaskevicius, T., Kanade, V., Rebeschini, P.: Implicit regularization for optimal sparse recovery. Advances in Neural Information Processing Systems32(2019)

2019
[30]

In: 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp

Wu, R., Ma, W.K., Fu, X.: A stochastic maximum-likelihood framework for simplex structured matrix factorization. In: 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2557–2561. IEEE (2017)

2017
[31]

In: 2019 27th European Signal Processing Conference (EUSIPCO), pp

Yeredor, A., Haardt, M.: Estimation of a low-rank probability-tensor from sample sub- tensors via joint factorization minimizing the Kullback-Leibler divergence. In: 2019 27th European Signal Processing Conference (EUSIPCO), pp. 1–5. IEEE (2019)

2019
[32]

arXiv preprint arXiv:1903.09367 2(4), 8 (2019) Title Suppressed Due to Excessive Length 23 A Appendix (Results with Different Learning Rates) Fig

Zhao, P., Yang, Y., He, Q.C.: Implicit regularization via Hadamard product over- parametrization in high-dimensional linear regression. arXiv preprint arXiv:1903.09367 2(4), 8 (2019) Title Suppressed Due to Excessive Length 23 A Appendix (Results with Different Learning Rates)...

1903 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

SIAM Journal on Mathematics of Data Science3(2), 593–623 (2021)

Abdolali, M., Gillis, N.: Simplex-structured matrix factorization: Sparsity-based iden- tifiability and provably correct algorithms. SIAM Journal on Mathematics of Data Science3(2), 593–623 (2021)

2021

[2] [2]

Princeton University Press, Princeton, NJ (2008)

Absil, P.A., Mahony, R., Sepulchre, R.: Optimization Algorithms on Matrix Manifolds. Princeton University Press, Princeton, NJ (2008). URLhttp://press.princeton.edu/ titles/8586.html

2008

[3] [3]

SIAM Journal on Optimization22(1), 135–158 (2012)

Absil, P.A., Malick, J.: Projection-like retractions on matrix manifolds. SIAM Journal on Optimization22(1), 135–158 (2012). DOI 10.1137/100802529. URLhttp://link. aip.org/link/?SJE/22/135/1

work page doi:10.1137/100802529 2012

[4] [4]

IEEE Transactions on Signal Processing69, 6195–6205 (2021)

Amiridi, M., Kargas, N., Sidiropoulos, N.D.: Information-theoretic feature selection via tensor decomposition and submodularity. IEEE Transactions on Signal Processing69, 6195–6205 (2021)

2021

[5] [5]

Society for Industrial and Applied Math- ematics (2019)

Arag´ on, F.J., et al.: Nonlinear Optimization. Society for Industrial and Applied Math- ematics (2019)

2019

[6] [6]

Journal of the Operational Research Society 48(3), 334–334 (1997)

Bertsekas, D.P.: Nonlinear programming. Journal of the Operational Research Society 48(3), 334–334 (1997)

1997

[7] [7]

SIAM Journal on Optimization10(4), 1196–1211 (2000)

Birgin, E.G., Mart ´ ınez, J.M., Raydan, M.: Nonmonotone spectral projected gradient methods on convex sets. SIAM Journal on Optimization10(4), 1196–1211 (2000)

2000

[8] [8]

Springer Science & Business Media (2006)

Bonnans, J.F., Gilbert, J.C., Lemar´ echal, C., Sagastiz´ abal, C.A.: Numerical optimiza- tion: theoretical and practical aspects. Springer Science & Business Media (2006)

2006