Pith. sign in

REVIEW 4 minor 300 references

Noisy SGD on wide two-layer ReLU nets collapses to at most 2P-1 effective directions set by the training data's linear dichotomies.

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

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

T0 review · grok-4.5

2026-07-12 01:11 UTC pith:LKR7LU7Y

load-bearing objection Solid multivariate extension of Shevchenko et al. that actually characterizes the stationary measure (width collapse + non-redundancy), with limitations already scoped correctly by the authors.

arxiv 2607.03613 v1 pith:LKR7LU7Y submitted 2026-07-03 cs.LG

Implicit Bias of SGD in Multivariate ReLU Networks: Effective Width Collapse

Shuang Liang , Tom Jacobs , Guido Mont\'ufar This is my paper
classification cs.LG
keywords implicit biasstochastic gradient descentmean-field limitReLU networksfeature learningeffective width collapsepiecewise affinelinear dichotomies
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Overparameterized networks can represent many interpolators, yet training with noisy SGD systematically picks particular ones. This paper studies that selection for multivariate regression with wide two-layer ReLU networks in the mean-field regime. It shows that, in the vanishing-noise limit, any accumulation point of the stationary measures yields a continuous piecewise-affine predictor whose kink hyperplanes are controlled by the training inputs: at most 2P-1 of them, where P counts the linear dichotomies realizable on those inputs. Simultaneously the input weights and biases of the infinite-width network align onto at most that many rays, so effective width collapses. Distinct rays induce incomparable ternary activation patterns on the data, so the learned features are non-redundant. The complexity of the selected predictor is therefore governed by the combinatorial geometry of the training set rather than by network width.

Core claim

Despite infinite overparameterization, noisy SGD (vanishing temperature, positive weight decay) selects continuous piecewise-affine predictors whose kink hyperplanes number at most 2P(X)-1 and whose input-weight/bias marginal is supported on at most that many rays; each ray induces a distinct, incomparable ternary activation pattern on the training inputs.

What carries the argument

The Boltzmann fixed-point characterization of the unique free-energy minimizer of the mean-field PDE, analyzed after a bounded smooth approximation of ReLU; the Quadratic Positivity Condition and the subsequent shrinkage of a data-induced cluster set on the sphere of input weights and biases determine both the vanishing Hessian outside finitely many hyperplanes and the support of the limiting measure.

Load-bearing premise

The joint infinite-width, vanishing-noise, ReLU limit is studied only through accumulation points of approximating sequences, whose existence is not proved for the measures themselves, and the limiting weight-decay strength must stay strictly positive.

What would settle it

Train a wide two-layer ReLU net by noisy SGD with vanishing noise and positive weight decay on a fixed multivariate data set; if the trained input weights/biases fail to concentrate on far fewer than 2P-1 rays, or if the learned function has more than 2P-1 kink hyperplanes that violate the ternary-pattern non-redundancy conditions, the claim is false.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 4 minor

Summary. The paper studies the implicit bias of noisy SGD for wide two-layer ReLU networks in multivariate regression under the mean-field regime of Mei et al. (2018). Training is approximated by a Wasserstein gradient flow of a free energy that admits a unique stationary measure satisfying a Boltzmann fixed-point condition. Working with bounded, smoothed approximations of ReLU and then taking limits, the authors characterize accumulation points of the stationary measures and the associated predictors as temperature vanishes and the truncation is removed. Theorem 4 shows that any such limiting predictor is continuous piecewise affine, with kink hyperplanes numbering at most 2P(X)-1, where P(X) is the number of linear dichotomies of the training inputs. Theorem 5 shows that the marginal of input weights and biases is supported on at most that many rays, yielding effective width collapse. Theorem 6 establishes a non-redundancy property: distinct rays induce incomparable ternary activation patterns on the training data. The proofs shift the analysis from input space to parameter space, introduce a Quadratic Positivity Condition and a cluster set on the sphere, and prove shrinkage of that set to finitely many points. Experiments illustrate alignment and piecewise affinity.

Significance. If correct, the results give the first derivation of effective width collapse as an implicit bias of noisy SGD for multivariate ReLU regression with essentially arbitrary finite training data, together with a data-dependent combinatorial bound controlled by P(X). This extends the univariate analysis of Shevchenko et al. (2022) both to higher input dimension and to a structural description of the stationary measure itself (not only the represented function). The bound is shown to improve on related static TV-norm bounds of de Dios–Bruna and Del Grande et al. for generic data. The geometric argument (cluster-set shrinkage on the sphere, non-redundancy of ternary patterns) is reusable and the appendices supply a detailed derivation chain. Code is provided. The main caveats—analysis of accumulation points rather than proven joint limits, and the requirement that limiting weight decay stay positive—are stated explicitly by the authors and do not undermine the claims as scoped.

minor comments (4)
  1. The existence of 2-Wasserstein convergent subsequences for the measures (needed for Theorem 5) is left more open than the pointwise existence for the predictors (Proposition 19). A short clarifying sentence in §4.2 or Appendix G would help readers see the precise gap.
  2. Figures 1–3 and 5–6 are informative but would benefit from a brief caption note on how the learned directions were extracted (DBSCAN thresholds, norm cutoff) so that the experimental protocol is fully self-contained.
  3. In the comparison with Del Grande et al. (Appendix F), the improvement factor 2^d/(d+1)^2 is stated for the regime 2d+1≤M; a one-line numerical example for small (d,M) would make the improvement more tangible.
  4. Notation for the cluster set (Ω̂^α vs. Ω̂*) and the sectors O_k is introduced gradually across §6; a short summary table or paragraph at the start of §6 would improve readability of the proof outline.

Circularity Check

0 steps flagged

No significant circularity: structural claims follow from the Boltzmann fixed-point of Mei et al. via independent geometric analysis of the cluster set; accumulation-point scoping is explicit and non-circular.

full rationale

The derivation chain begins from the free-energy minimizer and Boltzmann fixed-point condition of Mei et al. (2018, Theorem 1 / eq. 5), which is an external result under stated regularity assumptions on the (smoothed/truncated) activation. From there the paper derives a Hessian bound (Prop. 8), introduces the Quadratic Positivity Condition on the parameter-space sectors induced by the training inputs (Prop. 9), proves shrinkage of the associated cluster set ˆΩ* to at most 2P(X)-1 points on the sphere (Prop. 10, using only the classical count of 2P sectors from Cover 1965 and a merging argument that excludes the all-negative sector), and obtains vanishing Hessian outside the corresponding hyperplanes (Thm. 11) together with support of the input-weight/bias marginal on the rays (Thm. 5) and non-redundancy of ternary patterns (Prop. 13 / Thm. 6). None of these steps defines a quantity in terms of the final bound and recovers it, fits a free parameter to data and re-labels the fit as a prediction, or imports a uniqueness theorem from overlapping authors as an external fact that forces the claim. The restriction to accumulation points of sequences (β_n, λ_n, m_n) (rather than a proven joint limit m, β → ∞) is stated explicitly (end of §3.2, Assumption 2, Prop. 10, §5 comparison with Shevchenko et al.) and is accompanied by existence of such points (Prop. 19, Lemma 15); the geometric conclusions hold for every such accumulation point and do not rely on uniqueness of a joint limit. Self-citations are to technical lemmas of the univariate predecessor or to the authors’ own related work on other regimes; none is load-bearing for the central support/non-redundancy statements. P(X) is the classical combinatorial quantity of Cover (1965). Consequently the claimed effective-width collapse is a genuine structural consequence of the mean-field stationary condition, not a restatement of its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 2 invented entities

The paper rests on the mean-field convergence theorem of Mei et al. (2018) for noisy SGD, on standard optimal-transport and free-energy arguments, and on a technical smooth truncation of ReLU that is removed by double limits. No numerical free parameters are fitted; the only free constants are those appearing in the existence assumptions of the cited mean-field theory. Invented technical objects (cluster set, Quadratic Positivity Condition) are intermediate analytic devices, not new physical or statistical entities.

axioms (4)
  • domain assumption Noisy SGD with step-size scheme s_{k,N}=ε ξ(kε) converges, in the joint limit N→∞, ε→0, to the Wasserstein gradient flow of the free energy F_{β,λ}^σ whose unique minimizer satisfies the Boltzmann fixed-point equation (Mei et al. 2018, Thm. 1 / Lemma 10.3).
    Invoked as Theorem 1; supplies the unique stationary measure whose structure is analyzed.
  • ad hoc to paper The m-truncated, τ-smoothed ReLU neuron converges pointwise to the true ReLU as m,τ→∞ and satisfies the C^4 boundedness hypotheses of the mean-field theorem.
    Constructed in Appendix A following Shevchenko et al.; needed because raw ReLU violates the regularity assumptions of Mei et al.
  • ad hoc to paper Any sequence (β_n,λ_n,m_n) obeying Assumption 2 with m_n,β_n→∞ and λ_n→λ̄>0 admits a subsequence along which the network functions converge pointwise and (for Thm. 5) the measures converge in 2-Wasserstein distance.
    Existence of the joint limit is left open; the paper works exclusively with accumulation points (Prop. 19, Lemma 15).
  • standard math Standard facts on hyperplane arrangements, Cover’s bound on linear dichotomies, and the geometry of quadratic cones on the sphere.
    Used to count sectors (2P) and to prove that the limiting cluster set contains at most 2P-1 points.
invented entities (2)
  • Cluster set Ω̂^α (and its limit Ω̂*) no independent evidence
    purpose: Collects the spherical cross-sections of the non-positive sets of the quadratic forms that appear in the Hessian bound; its shrinkage controls both the vanishing of the Hessian and the support of the stationary measure.
    Purely analytic device introduced in §6.4; no independent empirical existence claimed.
  • Quadratic Positivity Condition no independent evidence
    purpose: A geometric criterion on the parameter space guaranteeing that the Hessian of the solution function vanishes in the zero-temperature / infinite-m limit at a given test point x.
    Technical hypothesis used to obtain the piecewise-affine conclusion; not an ontological claim.

pith-pipeline@v1.1.0-grok45 · 61992 in / 3033 out tokens · 29009 ms · 2026-07-12T01:11:06.962420+00:00 · methodology

0 comments
read the original abstract

We study the implicit bias of noisy stochastic gradient descent in training wide two-layer ReLU networks for multivariate regression. In a mean-field regime, the training dynamics are approximated by a Wasserstein gradient flow that converges to a unique stationary measure. We characterize the structure of this stationary measure and the predictor it represents. We show that, despite the network being infinitely overparameterized, the learned predictor admits an effectively finite representation: the input weights and biases align along finitely many directions, leading to an effective width collapse. In particular, the solution function is continuous piecewise affine, with affine regions determined by the cells of a finite hyperplane arrangement. The number of learned directions, and hence hyperplanes, is bounded above by $2\mathcal{P}-1$, where $\mathcal{P}$ denotes the number of linear dichotomies realizable on the training inputs. We further establish a non-redundancy property of the learned representation by proving that each learned direction induces a unique ternary activation pattern on the training data. Consequently, the complexity of the learned predictor is governed by the combinatorial geometry of the training data.

Figures

Figures reproduced from arXiv: 2607.03613 by Guido Mont\'ufar, Shuang Liang, Tom Jacobs.

Figure 1
Figure 1. Figure 1: A two-layer ReLU network with N = 1000 hidden neurons and two input dimen￾sions, trained via SGD with weight decay. Left: the learned function is piecewise affine and approximately fits the training data (blue dots). Middle: the training trajectories of the input weights and biases (w i 1 , wi 2 , bi ) ∈ R 3 , i = 1, . . . , N align along a small number of directions, and the effective width of the network… view at source ↗
Figure 2
Figure 2. Figure 2: Non-redundancy of the learned features. Left: input points and kink hyperplanes of the trained network under the same setting of [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A two-layer univariate ReLU network with N = 1000 hidden neurons, trained via SGD with weight decay on a symmetric data set of size two. Blue dots represent the training data and orange curves the learned functions. The figure illustrates how the location of the kink hyperplanes (knots) is determined by the training data, as predicted in Proposition 7. contains at most two elements. Moreover, if D(Aˆ) has … view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the Quadratic Positivity Condition for d = 2. For a fixed k, the red antipodal disks represent Ωˆα k , the non-positive set of the quadratic function f α k ; the region bounded by the blue arcs is Oˆ k, the sector corresponding to a particular binary pattern on the training inputs; and the green arc is Gˆ x ∩ Oˆ k, the set of parameters defining a hyperplane that passes through x while indu… view at source ↗
Figure 5
Figure 5. Figure 5: A two-layer ReLU network is trained via SGD without noise on two-dimensional input data, under various levels of weight decay λ. The figure shows that the trained function is piecewise affine for all values of λ, whereby larger λ promotes a smaller Lipschitz constant [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A two-layer ReLU network is trained via noisy SGD with weight decay on two￾dimensional input data, under various noise levels β −1 . The figure shows that, as the noise level tends to zero, the trained function converges to a piecewise affine function, while for larger noise levels, the function is increasingly smooth. data set via SGD for 106 iterations. The network width, initialization, and baseline SGD… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

300 extracted references · 4 canonical work pages

  1. [1]

    Advances in neural information processing systems , volume=

    The rich and the simple: On the implicit bias of adam and sgd , author=. Advances in neural information processing systems , volume=

  2. [2]

    Proceedings of the Second International Conference on Knowledge Discovery and Data Mining , pages=

    A density-based algorithm for discovering clusters in large spatial databases with noise , author=. Proceedings of the Second International Conference on Knowledge Discovery and Data Mining , pages=

  3. [3]

    The Fourteenth International Conference on Learning Representations , year=

    Never Saddle for Reparameterized Steepest Descent as Mirror Flow , author=. The Fourteenth International Conference on Learning Representations , year=

  4. [4]

    International Conference on Learning Representations , volume=

    Flavors of margin: Implicit bias of steepest descent in homogeneous neural networks , author=. International Conference on Learning Representations , volume=

  5. [5]

    2026 , eprint=

    Implicit Bias of Mirror Flow in Homogeneous Neural Networks: Sparse and Dense Feature Learning , author=. 2026 , eprint=

    2026

  6. [6]

    On sparsity in overparametrised shallow

    de Dios, Jaume and Bruna, Joan , journal=. On sparsity in overparametrised shallow

  7. [7]

    Global Minimizers of _p -Regularized Objectives Yield the Sparsest

    Julia B Nakhleh and Robert D Nowak , booktitle=. Global Minimizers of _p -Regularized Objectives Yield the Sparsest. 2026 , url=

    2026

  8. [8]

    Journal of Machine Learning Research , volume=

    Early alignment in two-layer networks training is a two-edged sword , author=. Journal of Machine Learning Research , volume=

  9. [9]

    Advances in Neural Information Processing Systems , volume=

    Towards understanding the condensation of neural networks at initial training , author=. Advances in Neural Information Processing Systems , volume=

  10. [10]

    arXiv preprint arXiv:2009.10713 , year=

    Towards a mathematical understanding of neural network-based machine learning: what we know and what we don't , author=. arXiv preprint arXiv:2009.10713 , year=

    Pith/arXiv arXiv 2009

  11. [11]

    Tensor programs

    Yang, Greg and Hu, Edward J , booktitle=. Tensor programs. 2021 , organization=

    2021

  12. [12]

    arXiv preprint arXiv:2011.14522 , year=

    Feature learning in infinite-width neural networks , author=. arXiv preprint arXiv:2011.14522 , year=

    Pith/arXiv arXiv 2011

  13. [13]

    Implicit Compressibility of Overparametrized Neural Networks Trained with Heavy-Tailed

    Yijun Wan and Melih Barsbey and Abdellatif Zaidi and Umut Simsekli , booktitle=. Implicit Compressibility of Overparametrized Neural Networks Trained with Heavy-Tailed. 2024 , url=

    2024

  14. [14]

    Journal of Machine Learning Research , volume=

    Law of large numbers and central limit theorem for wide two-layer neural networks: the mini-batch and noisy case , author=. Journal of Machine Learning Research , volume=

  15. [15]

    SIAM Journal on Applied Mathematics , volume=

    Mean field analysis of neural networks: A law of large numbers , author=. SIAM Journal on Applied Mathematics , volume=. 2020 , publisher=

    2020

  16. [16]

    Conference on learning theory , pages=

    Benefits of depth in neural networks , author=. Conference on learning theory , pages=. 2016 , organization=

    2016

  17. [17]

    Rademacher and

    Bartlett, Peter L and Mendelson, Shahar , journal=. Rademacher and

  18. [18]

    2026 , eprint=

    Better Neural Network Expressivity: Subdividing the Simplex , author=. 2026 , eprint=

    2026

  19. [19]

    SIAM Journal on Discrete Mathematics , volume =

    Hertrich, Christoph and Basu, Amitabh and Di Summa, Marco and Skutella, Martin , title =. SIAM Journal on Discrete Mathematics , volume =. 2023 , nodoi =. https://doi.org/10.1137/22M1489332 , abstract =

    doi:10.1137/22m1489332 2023

  20. [20]

    2020 , cdate=

    Henning Petzka and Martin Trimmel and Cristian Sminchisescu , title=. 2020 , cdate=

    2020

  21. [21]

    A Complete Symmetry Classification of Shallow

    Pranavkrishnan Ramakrishnan , year=. A Complete Symmetry Classification of Shallow. 2604.14037 , archivePrefix=

    Pith/arXiv arXiv

  22. [22]

    The Symmetries of Three-Layer

    Johanna Marie Gegenfurtner and Moritz Grillo and Guido Montúfar , year=. The Symmetries of Three-Layer. 2605.18319 , archivePrefix=

    Pith/arXiv arXiv

  23. [23]

    Lampert , booktitle=

    Mary Phuong and Christoph H. Lampert , booktitle=. Functional vs.\ parametric equivalence of. 2020 , url=

    2020

  24. [24]

    Hidden Symmetries of

    Grigsby, Elisenda and Lindsey, Kathryn and Rolnick, David , booktitle =. Hidden Symmetries of. 2023 , noeditor =

    2023

  25. [25]

    Moritz Grillo and Guido Montúfar , year=. Most. 2605.03601 , archivePrefix=

    Pith/arXiv arXiv

  26. [26]

    2005 , publisher=

    Gradient flows: in metric spaces and in the space of probability measures , author=. 2005 , publisher=

    2005

  27. [27]

    Journal of Machine Learning Research , year =

    Tolga Ergen and Mert Pilanci , title =. Journal of Machine Learning Research , year =

  28. [28]

    Proceedings of the 37th International Conference on Machine Learning , pages =

    Neural Networks are Convex Regularizers: Exact Polynomial-time Convex Optimization Formulations for Two-layer Networks , author =. Proceedings of the 37th International Conference on Machine Learning , pages =. 2020 , editor =

    2020

  29. [29]

    Proceedings of the 58th Annual ACM Symposium on Theory of Computing , pages =

    Bakaev, Egor and Brunck, Florestan and Hertrich, Christoph and Stade, Jack and Yehudayoff, Amir , title =. Proceedings of the 58th Annual ACM Symposium on Theory of Computing , pages =. 2026 , noisbn =

    2026

  30. [30]

    Journal of Machine Learning Research , year =

    Tao Luo and Zhi-Qin John Xu and Zheng Ma and Yaoyu Zhang , title =. Journal of Machine Learning Research , year =

  31. [31]

    1975 , publisher=

    Facing up to Arrangements: Face-Count Formulas for Partitions of Space by Hyperplanes , author=. 1975 , publisher=

    1975

  32. [32]

    How Does the

    Lai, Kuo-Wei and Wang, Guanghui and Tao, Molei and Muthukumar, Vidya , journal=. How Does the

  33. [33]

    The Fourteenth International Conference on Learning Representations , year=

    Gradient Descent with Large Step Sizes: Chaos and Fractal Convergence Region , author=. The Fourteenth International Conference on Learning Representations , year=

  34. [34]

    Advances in neural information processing systems , volume=

    Implicit regularization in deep learning may not be explainable by norms , author=. Advances in neural information processing systems , volume=

  35. [35]

    Advances in neural information processing systems , volume=

    Convex neural networks , author=. Advances in neural information processing systems , volume=

  36. [36]

    Journal of Machine Learning Research , volume=

    Breaking the curse of dimensionality with convex neural networks , author=. Journal of Machine Learning Research , volume=

  37. [37]

    SIAM journal on mathematical analysis , volume=

    The variational formulation of the Fokker--Planck equation , author=. SIAM journal on mathematical analysis , volume=. 1998 , publisher=

    1998

  38. [38]

    Annales de l'Institut Henri Poincare (B) Probabilites et statistiques , volume=

    Uniform-in-time propagation of chaos for mean field Langevin dynamics , author=. Annales de l'Institut Henri Poincare (B) Probabilites et statistiques , volume=. 2025 , organization=

    2025

  39. [39]

    arXiv preprint arXiv:2603.01842 , year=

    Uniform-in-time concentration in two-layer neural networks via transportation inequalities , author=. arXiv preprint arXiv:2603.01842 , year=

    arXiv

  40. [40]

    Nearly-tight

    Bartlett, Peter L and Harvey, Nick and Liaw, Christopher and Mehrabian, Abbas , journal=. Nearly-tight

  41. [41]

    1993 , publisher=

    Topologies on closed and closed convex sets , author=. 1993 , publisher=

    1993

  42. [42]

    Geometric combinatorics , volume=

    An introduction to hyperplane arrangements , author=. Geometric combinatorics , volume=

  43. [43]

    The annals of mathematical statistics , pages=

    A stochastic approximation method , author=. The annals of mathematical statistics , pages=. 1951 , publisher=

    1951

  44. [44]

    Annales de l'Institut Henri Poincare (B) Probabilites et statistiques , volume=

    Mean-field Langevin dynamics and energy landscape of neural networks , author=. Annales de l'Institut Henri Poincare (B) Probabilites et statistiques , volume=. 2021 , organization=

    2021

  45. [45]

    arXiv preprint arXiv:2603.17785 , year=

    A Dual Certificate Approach to Sparsity in Infinite-Width Shallow Neural Networks , author=. arXiv preprint arXiv:2603.17785 , year=

    arXiv

  46. [46]

    IEEE Transactions on Electronic Computers , volume=

    Geometrical and Statistical Properties of Systems of Linear Inequalities with Applications in Pattern Recognition , author=. IEEE Transactions on Electronic Computers , volume=

  47. [47]

    2013 , publisher=

    Convergence of probability measures , author=. 2013 , publisher=

    2013

  48. [48]

    Handbook of differential equations: evolutionary equations , volume=

    Gradient flows of probability measures , author=. Handbook of differential equations: evolutionary equations , volume=. 2007 , publisher=

    2007

  49. [49]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Deep linear networks for regression are implicitly regularized towards flat minima , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  50. [50]

    The Thirteenth International Conference on Learning Representations , year=

    Universal Sharpness Dynamics in Neural Network Training: Fixed Point Analysis, Edge of Stability, and Route to Chaos , author=. The Thirteenth International Conference on Learning Representations , year=

  51. [51]

    Ussr computational mathematics and mathematical physics , volume=

    Some methods of speeding up the convergence of iteration methods , author=. Ussr computational mathematics and mathematical physics , volume=. 1964 , publisher=

    1964

  52. [52]

    Learning multiple layers of features from tiny images.(2009) , author=

    2009

  53. [53]

    Advances in neural information processing systems , volume=

    Stochasticity of deterministic gradient descent: Large learning rate for multiscale objective function , author=. Advances in neural information processing systems , volume=

  54. [54]

    2003 , publisher=

    The symmetry perspective: from equilibrium to chaos in phase space and physical space , author=. 2003 , publisher=

    2003

  55. [55]

    Physics Letters A , volume=

    Final state sensitivity: an obstruction to predictability , author=. Physics Letters A , volume=. 1983 , publisher=

    1983

  56. [56]

    Mathematische Zeitschrift , volume=

    Geometry of real polynomial mappings , author=. Mathematische Zeitschrift , volume=. 2002 , publisher=

    2002

  57. [57]

    2000 , publisher=

    Introduction to topological manifolds , author=. 2000 , publisher=

    2000

  58. [58]

    Physical Review E , volume=

    Mixed basin boundary structures of chaotic systems , author=. Physical Review E , volume=. 1999 , publisher=

    1999

  59. [59]

    Physical review letters , volume=

    Sporadically fractal basin boundaries of chaotic systems , author=. Physical review letters , volume=. 1999 , publisher=

    1999

  60. [60]

    Reviews of Modern Physics , volume=

    Fractal structures in nonlinear dynamics , author=. Reviews of Modern Physics , volume=. 2009 , publisher=

    2009

  61. [61]

    Physical Review Letters , volume=

    Fractal basin boundaries, long-lived chaotic transients, and unstable-unstable pair bifurcation , author=. Physical Review Letters , volume=. 1983 , publisher=

    1983

  62. [62]

    Physica D: Nonlinear Phenomena , volume=

    Wada basin boundaries and basin cells , author=. Physica D: Nonlinear Phenomena , volume=. 1996 , publisher=

    1996

  63. [63]

    Physica D: Nonlinear Phenomena , volume=

    Basins of wada , author=. Physica D: Nonlinear Phenomena , volume=. 1991 , publisher=

    1991

  64. [64]

    Wada basins and chaotic invariant sets in the H

    Aguirre, Jacobo and Vallejo, Juan C and Sanju. Wada basins and chaotic invariant sets in the H. Physical Review E , volume=. 2001 , publisher=

    2001

  65. [65]

    Proceedings of the IEEE international conference on computer vision , pages=

    Unifying nuclear norm and bilinear factorization approaches for low-rank matrix decomposition , author=. Proceedings of the IEEE international conference on computer vision , pages=

  66. [66]

    Information and Inference: A Journal of the IMA , volume=

    The non-convex geometry of low-rank matrix optimization , author=. Information and Inference: A Journal of the IMA , volume=. 2019 , publisher=

    2019

  67. [67]

    Sur les transformations ponctuelles , url =

    Hadamard , journal =. Sur les transformations ponctuelles , url =

  68. [68]

    The American Mathematical Monthly , volume=

    Period three implies chaos , author=. The American Mathematical Monthly , volume=. 1975 , publisher=

    1975

  69. [69]

    1995 , publisher=

    Introduction to the modern theory of dynamical systems , author=. 1995 , publisher=

    1995

  70. [70]

    Pure and Applied Geophysics , volume=

    The essence of chaos , author=. Pure and Applied Geophysics , volume=. 1996 , publisher=

    1996

  71. [71]

    A condition for the properness of polynomial maps , journal =

    Ngu. A condition for the properness of polynomial maps , journal =. 2009 , pages =

    2009

  72. [72]

    Geometry of real polynomial mappings , Ty =

    Jelonek, Zbigniew , Da =. Geometry of real polynomial mappings , Ty =. Mathematische Zeitschrift , Number =. 2002 , Bdsk-Url-1 =. doi:10.1007/s002090100298 , Id =

    work page doi:10.1007/s002090100298 2002

  73. [73]

    Properness of Polynomial Maps with

    Fukui, Toshizumi and Tsuchiya, Takeki , Da =. Properness of Polynomial Maps with. Arnold Mathematical Journal , Number =. 2023 , Bdsk-Url-1 =. doi:10.1007/s40598-022-00205-2 , Id =

    work page doi:10.1007/s40598-022-00205-2 2023

  74. [74]

    Testing sets for properness of polynomial mappings , Ty =

    Jelonek, Zbigniew , Da =. Testing sets for properness of polynomial mappings , Ty =. Mathematische Annalen , Number =. 1999 , Bdsk-Url-1 =. doi:10.1007/s002080050316 , Id =

    work page doi:10.1007/s002080050316 1999

  75. [75]

    2008 , publisher=

    Equilibrium States and the Ergodic Theory of Anosov Diffeomorphisms , author=. 2008 , publisher=

    2008

  76. [76]

    , number=

    Anosov, D.V. , number=. Geodesic Flows on Closed. 1969 , publisher=

    1969

  77. [77]

    Smale , title =

    S. Smale , title =. Bulletin of the American Mathematical Society , number =

  78. [78]

    2014 , publisher=

    Topological dynamical systems: an introduction to the dynamics of continuous mappings , author=. 2014 , publisher=

    2014

  79. [79]

    2007 , publisher=

    Discrete chaos: with applications in science and engineering , author=. 2007 , publisher=

    2007

  80. [80]

    2012 , publisher=

    One-dimensional dynamics , author=. 2012 , publisher=

    2012

Showing first 80 references.