Pith. sign in

REVIEW 3 major objections 4 minor 70 references

Impact of Bottleneck Layers and Skip Connections on the Generalization of Linear Denoising Autoencoders

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Bottleneck width creates a bias-variance trade-off in linear denoising autoencoders, and skip connections damp the overparameterization peak, according to closed-form test-risk formulas.

desk verdict A real step forward for linear DAE theory, but the skip-connection risk formula in Theorem 3.5 has a spurious 1/c that must be fixed. read the letter →

arxiv 2505.24668 v1 pith:KF6JB2ZF submitted 2025-05-30 stat.ML cs.LG

classification stat.MLcs.LG MSC 60B2062H2568T07
keywords denoisingautoencodersbottlenecklayersskipconnectionsgeneralizationdoubledescentbias-variancetrade-offlinearneuralnetworksrandommatrixtheory
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

The paper derives closed-form test-risk formulas for two-layer linear denoising autoencoders with a bottleneck layer, both with and without a skip connection, in the overparameterized regime $d \geq n+r$. It shows that the global minimizer is $W_* = P_{[k]}(X)(X+A)^\dagger$ and that the bottleneck width $k$ acts as a complexity dial: increasing $k$ reduces bias but increases variance, creating a modern bias-variance trade-off. It further shows that a skip connection multiplies the dangerous $(c-1)^{-1}$ variance term by an extra $1/d$ factor, flattening the peak near $c=d/n \approx 1$. A random-matrix analysis traces this flattening to weaker alignment between the noise singular vectors and the corrupted-data eigenvectors. A sympathetic reader would care because this gives the first analytic picture of how two architectural staples, bottlenecks and skip connections, shape generalization in an unsupervised setting.

What carries the argument

The load-bearing object is the closed-form characterization of all critical points of the product-regularized two-layer linear loss: every critical point has the form $\hat W_c = U_{G,I}U_{G,I}^\top YZ^\top(\tilde Z\tilde Z^\top)^{-1}$, and in the ridgeless limit $W_c = P_I(Y)Z^\dagger$, with the global minimizer $W_* = P_{[k]}(X)(X+A)^\dagger$. The risk proofs expand $(X+A)^\dagger$ through a pseudo-inverse perturbation formula and then replace expectations of products of the random matrices $D(P^\top P)^{-1}D$, $K_1^{-1}$, $ZD^{-1}$, and $HH^\top$ by products of expectations, relying on element-wise variances of order $O(1/d)$. For the skip-connection comparison, the paper introduces the rank-$1$ additive model $S=XX^\top+AA^\top$ and proves the alignment ratio identity $E[\langle u_i^A,u_j^S\rangle^2 / \langle u_1,u_j^S\rangle^2] = \Theta\left(\frac{1}{d(\lambda_i^A-\lambda_j^S)^2}\right)$, which is what makes the noise-alignment contribution to the variance parametrically smaller.

What would settle it

Fix a deterministic rank-$r$ signal with well-conditioned spectrum, draw Gaussian training and test noise with variances $\eta_{\mathrm{trn}}^2/d$ and $\eta_{\mathrm{tst}}^2/d$, and train both models to their ridgeless global minimizers. Compare the measured test risk with Eqs. (9) and (10) as $d$ grows with $c$ fixed near $1$: the formulas are falsified if the rescaled variance term $d(c-1)R$ does not converge to $\eta_{\mathrm{tst}}^2 \sum_{j\in[k]} \alpha_j^2/(1+\alpha_j^2)$, or if the skip-connected variance is not smaller by an extra factor of $1/d$.

Watch

Extended reading notes

Core claim

Under a deterministic low-rank signal model with Gaussian noise of variances $\eta_{\mathrm{trn}}^2/d$ and $\eta_{\mathrm{tst}}^2/d$, and with $d \geq n+r$, the paper claims exact control of the test risk of every critical point of the two-layer linear DAE. For the global minimizer $W_* = P_{[k]}(X)(X+A)^\dagger$ without a skip connection, the test risk is $R(W_*, X_{\mathrm{tst}}) = \frac{1}{N_{\mathrm{tst}}}\mathrm{Tr}(JLL^\top) + \frac{\eta_{\mathrm{tst}}^2 c}{d(c-1)}\sum_{j=1}^k \frac{\alpha_j^2}{1+\alpha_j^2} + o(1/d)$, where the diagonal entries of $J$ are $(1+\alpha_i^2)^{-2}$ for the selected directions and $1$ otherwise. The first term falls as $k$ grows, while the second rises, giving the bottleneck width a genuine bias-variance trade-off. With a skip connection, whose training objective is equivalent to predicting the noise $-A$ rather than the clean signal, the paper derives the corresponding risk formula whose variance term is smaller by an extra factor of $1/d$, so the $(c-1)^{-1}$ peak is damped, most visibly in the mildly overparameterized regime. The mechanism is identified in a rank-$1$ additive model $S = XX^\top + AA^\top$: the alignment between singular vectors of the noise and of $S$ is $\Theta(1/d)$ weaker than the alignment between singular vectors of the signal and of $S$, which suppresses the dominant small-eigenvalue contribution to the variance.

Load-bearing premise

The risk formulas rest on the assumption that the random matrices appearing inside the trace concentrate quickly enough that the expectation of a product equals the product of expectations up to $o(1)$ noise; if that joint concentration fails, the closed-form curves could acquire extra terms.

Editorial extensions

If this is right

  • For fixed input dimension $d$ and sample size $n$, the test risk as a function of bottleneck width $k$ is U-shaped: larger $k$ lowers bias but raises variance, and the optimal $k$ depends on the overparameterization ratio $c=d/n$.
  • Double descent in $d/n$ persists even when the bottleneck is small: the variance term diverges like $(c-1)^{-1}$ as $c$ approaches $1$, so the peak is a genuine feature of the architecture, not of unlimited rank.
  • Adding a skip connection changes the effective target from the clean signal to the noise, and the resulting variance term carries an extra $1/d$ factor; this dampens the peak near $c\approx 1$ without eliminating it.
  • In the skip-connected model, different critical points have nearly the same test error as the global minimizer, whereas in the bottleneck-only model different critical points can differ substantially in both bias and variance.
  • The variance reduction is traced to singular-vector alignment: noise singular vectors align with corrupted-data eigenvectors a factor $\Theta(1/d)$ more weakly than signal singular vectors do, and this suppresses the contribution of the smallest eigenvalues that dominate the variance.

Reading between the lines

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

  • Editorial inference, beyond the paper: the same concentration-plus-pseudo-inverse pipeline should apply to other minimum-norm interpolating autoencoders, such as masked autoencoders where the mask plays the role of the additive noise; the paper does not discuss this extension.
  • Editorial inference, beyond the paper: the paper's own remark about a four-layer linear model with the skip connection between hidden layers suggests a direct test of whether the constant $\eta_{\mathrm{tst}}^2$ bias can be removed while preserving the variance damping, since the two-layer input-output skip forces the network to predict full-rank noise.
  • Editorial inference, beyond the paper: the theory predicts that the optimal bottleneck width for fixed $n$ should grow with $d$ in the mildly overparameterized regime; a grid search over $k$ and $d$ on low-rank data would provide a sharp quantitative check of the predicted U-shape.
  • Editorial inference, beyond the paper: the alignment mechanism likely transfers to nonlinear denoisers trained with denoising score matching, where the effective target is also noise; the paper's linear analysis gives a concrete baseline for how much variance reduction one can attribute to the skip connection alone.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This paper studies two-layer linear denoising autoencoders trained by gradient flow on the squared reconstruction loss with product regularization, in the overparameterized regime d ≥ n + r. For a low-rank Gaussian noise data model (Assumption 3.1), it derives closed-form expressions for all critical points (Theorem 2.1) and specializes them to DAEs with and without an identity skip connection (Corollaries 2.4 and 2.5). It then states asymptotic test-risk formulas (Theorems 3.2 and 3.5) and decomposes them into bias and variance terms, arguing that the bottleneck width k induces a bias-variance trade-off and that skip connections damp the variance peak near c = d/n ≈ 1. Section 4 uses random matrix theory on a rank-1 additive model to explain the variance suppression via weaker eigenvector alignment, and CIFAR-10 experiments support the qualitative curves.

Significance. The paper makes a useful step beyond single-layer analyses: it gives a complete critical-point characterization for two-layer linear DAEs with bottleneck and skip connection, identifies the minimum-norm global minimizer, and produces interpretable closed-form risk formulas. The bias-variance decomposition of Eq. (9) and the variance-damping comparison between Eq. (9) and Eq. (10) are conceptually valuable, and the numerical experiments appear to match the intended predictions. However, the main theorem for the skip-connection model, Theorem 3.5, contains an algebraic error: the stated J_sc is inconsistent with the paper's own Lemma E.13 and with the mean computations in its proof. Because Eq. (10) is advertised in the abstract as one of the two central risk formulas, the paper cannot be accepted without correcting this formula and re-checking the figures that plot it.

major comments (3)
  1. [Theorem 3.5, Eq. (10), and Lemma E.13] The stated coefficient J_sc does not follow from the paper's own derivation. Lemma E.13 states E∥W_sc X_tst∥²_F = |I_a|/d Tr(((c−1)D²+I_d)(I_r+η_{trn}^{-2}D²)^{-2}LL^T)+O(d^{-1}); substituting this into the decomposition in Eq. (31) yields the bias term |I_a|/(dN_tst) Tr(J_correct LL^T) with J_correct,ii = ((c−1)σ_i²+1)/(1+η_{trn}^{-2}σ_i²)^2. Theorem 3.5 instead has J_sc,ii = (c+(c−1)σ_i²)/(c(1+η_{trn}^{-2}σ_i²)^2), which differs by a factor 1/c on the σ_i² term. The proof of Lemma E.13 also ends with a third, different form ((c−1)D²+cI_d)(cI_r+cη^{-2}D²)^{-2}. The theorem as stated is therefore not established by the preceding argument; please correct Eq. (10), align the closing display of Lemma E.13 with its statement, and verify that the theoretical curves in Figure 3 use the corrected formula.
  2. [Appendix E.1, Eqs. (28)-(30) and Lemmas E.2-E.13] The central derivations rely on replacing expectations of products of random matrix factors by products of expectations, justified by element-wise variance O(d^{-1}) for each factor. For products of four dependent factors such as D(P^T P)^{-1}D, K_1^{-1}, ZD^{-1}, and HH^T, the error involves joint fourth-order cumulants, not just pairwise variances. The manuscript cites [27] and [11] but does not state or prove a uniform lemma controlling these cross-terms. Since the closed-form risk formulas depend on this concentration step, please provide the missing concentration lemma or an explicit reference with the exact statement used.
  3. [Section 4, Definition 4.1 and Theorem 4.3] The alignment-suppression theorem is proved for the additive model S=XX^T+AA^T, whereas the DAE actually corresponds to the information-plus-noise model (X+A)(X+A)^T. Remark 4.2 only matches first moments of the two models; it does not imply that eigenvector alignment properties transfer. The paper acknowledges the simplification, but as written the 'origin' of the variance discrepancy is a heuristic supported by a related model, not a theorem for the actual model. Please either prove the analogous statement for (X+A)(X+A)^T (or a precise asymptotic equivalence) or explicitly label this part as a heuristic.
minor comments (4)
  1. [Appendix E.2, Definition E.14] Defining bias as the component that decreases with k and variance as the component that increases with k makes the bias-variance trade-off true by definition. The concrete decompositions in Eq. (28), Eq. (31), Lemmas E.3/E.4, and E.11/E.12 are the substantive statements; please present Definition E.14 as an interpretive summary rather than a formal definition.
  2. [Remark F.4] The case j=1 is dismissed 'for brevity' in a proof of a theorem about alignment; since the largest eigenvalue can dominate variance, this case should be either proved or explicitly excluded from the theorem statement.
  3. [Appendix E, notation] In Appendix E, D is used both as a d×d matrix and as an r×r block, and Lemma E.13 mixes I_d and I_r; please make the dimensions of all identity matrices and diagonal blocks explicit.
  4. [Theorems 3.2 and 3.5] The theorems state asymptotic o(1/d) and O(1/(dN_tst)) remainders while the text calls the setting 'non-asymptotic'; please clarify whether these are statements about sequences with d,n→∞ or finite-d bounds with explicit constants.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the risk formulas are derived from explicit critical-point expressions and concentration lemmas, not from fitted parameters or self-citations.

full rationale

The paper's derivation chain is self-contained in the sense relevant to circularity analysis. The closed-form risk formulas in Theorems 3.2 and 3.5 are obtained by substituting the explicitly characterized critical points W_c = P_Ix(X)(X+A)^+ and W^sc_c = -P_Ia(A)(X+A)^+ into the test-risk definitions, expanding the pseudo-inverse via the Woodbury-type identity from [45], and then applying concentration arguments (largely inherited from [27] and [38]) to replace expectations of products of random matrices by products of expectations. No parameter is fitted to data, no 'prediction' is defined in terms of a fitted value, and the formulas are not assumed as ansatze. The prior results [27, 38] are by different authors, so there is no load-bearing self-citation chain. The only passage that could raise a definitional concern is Definition E.14, which defines bias as the component that decreases with bottleneck width k and variance as the component that increases with k; this is tautological as a labeling device, but it is applied after the monotonicity of the terms in Eq. (9) and Eq. (10) is already established, and it does not enter the derivation of the risk formulas. The skeptic's suspected 1/c inconsistency in Lemma E.13/Theorem 3.5 is a mathematical correctness or proof-consistency issue, not a circularity: Eq. (10) is not equivalent to its inputs by construction, and correcting an algebraic factor would not change the derivation structure. Accordingly, the paper earns a 0 circularity score.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The central claim is derived analytically with no fitted parameters; k, η_trn, η_tst and the singular values of X are inputs from the data model. The results rest on the low-rank Gaussian data assumptions (Assumption 3.1), generic eigengap assumptions, the full-rank property of X+A, and standard random matrix concentration tools. No new entities are introduced.

assumptions (6)
  • domain assumption The clean data X is a deterministic low-rank matrix with rank r ≪ d,n, spectral norm Θ(1), and well-conditioned nonzero singular values; test data lies in the same subspace.
    Assumption 3.1. This data model enables the low-rank-plus-noise analysis; it is inherited from [27].
  • domain assumption The additive noise matrices A and A_tst have i.i.d. Gaussian entries with variance η²_trn/d and η²_tst/d.
    Assumption 3.1(3). Needed for Marchenko-Pastur and concentration arguments.
  • domain assumption X+A has full column rank n (holds almost surely when d ≥ n).
    Stated before Corollary 2.4 in Section 2.3; used to apply the pseudo-inverse expansion results.
  • domain assumption The nonzero eigenvalues of XX^T and AA^T have multiplicity one.
    Section 2.3; generic for random noise, used to enumerate critical points by ordered index sets.
  • domain assumption Gradient flow converges to the global minimizer of the regularized loss; the λ→0 ridgeless limit selects the minimum-norm solution.
    Section 2.1 and Theorem D.3. The paper proves the global minimizer is unique and that all other critical points are saddles, but convergence of the dynamics to this point is assumed.
  • standard math Standard random matrix theory results hold, including the Marchenko-Pastur law, concentration of quadratic forms, and resolvent identities.
    Used throughout Appendices C-F (Theorem C.2, Lemma C.4, etc.).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Impact of Bottleneck Layers and Skip Connections on the Generalization of Linear Denoising Autoencoders." pith.science (2026). https://pith.science/paper/KF6JB2ZF

@misc{pith2026250524668,
  author       = {Pith},
  title        = {Pith review of: Impact of Bottleneck Layers and Skip Connections on the Generalization of Linear Denoising Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KF6JB2ZF}},
  note         = {Machine review of arXiv:2505.24668}
}
read the original abstract

Modern deep neural networks exhibit strong generalization even in highly overparameterized regimes. Significant progress has been made to understand this phenomenon in the context of supervised learning, but for unsupervised tasks such as denoising, several open questions remain. While some recent works have successfully characterized the test error of the linear denoising problem, they are limited to linear models (one-layer network). In this work, we focus on two-layer linear denoising autoencoders trained under gradient flow, incorporating two key ingredients of modern deep learning architectures: A low-dimensional bottleneck layer that effectively enforces a rank constraint on the learned solution, as well as the possibility of a skip connection that bypasses the bottleneck. We derive closed-form expressions for all critical points of this model under product regularization, and in particular describe its global minimizer under the minimum-norm principle. From there, we derive the test risk formula in the overparameterized regime, both for models with and without skip connections. Our analysis reveals two interesting phenomena: Firstly, the bottleneck layer introduces an additional complexity measure akin to the classical bias-variance trade-off -- increasing the bottleneck width reduces bias but introduces variance, and vice versa. Secondly, skip connection can mitigate the variance in denoising autoencoders -- especially when the model is mildly overparameterized. We further analyze the impact of skip connections in denoising autoencoder using random matrix theory and support our claims with numerical evidence.

Figures

Figures reproduced from arXiv: 2505.24668 by the authors.

Figure 2
Figure 2. Effect of Bottleneck. Test errors on CIFAR-10 illustrating how the bottleneck dimension k influences generalization. The left plot corresponds to the model without skip connections. The center shows results for the model with skip connections. The right subfigure is constructed by jointly increasing both k and d, using the corresponding test errors from the left plot. As seen in the left and center plots, the optima… view at source ↗
Figure 3
Figure 3. Effect of Skip Connections. Experiments on CIFAR-10 (Data rank is fixed at r = 100). Solid lines represent theoretical predictions, while cross markers indicate empirical results. Dotted lines correspond to empirical values in the underparameterized regime (solutions derived from [8]). Red lines and markers denote results for the model with skip connections. The left subfigure shows the test error; the center subfig… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 53 canonical work pages

  1. [27]

    Double Descent and Overfitting under Noisy Inputs and Distribution Shift for Linear Denoisers

    Chinmaya Kausik, Kashvi Srivastava, and Rishi Sonthalia. Double Descent and Overfitting under Noisy Inputs and Distribution Shift for Linear Denoisers. 2024. arXiv: 2305.17297 [cs.LG]. URL: https://arxiv.org/abs/2305.17297

  2. [11]

    On the Exact Covariance of Products of Random Variables

    George W. Bohrnstedt and Arthur S. Goldberger. “On the Exact Covariance of Products of Random Variables”. In: Journal of the American Statistical Association 64.328 (1969), pp. 1439–1442. ISSN : 01621459, 1537274X. URL: http : / / www . jstor . org / stable / 2286081 (visited on 08/10/2024)

  3. [1]

    The loss landscape of deep linear neural networks: a second-order analysis

    El Mehdi Achour, François Malgouyres, and Sébastien Gerchinovitz. The loss landscape of deep linear neural networks: a second-order analysis. 2024. arXiv: 2107.13289 [math.ST]. URL: https://arxiv.org/abs/2107.13289

  4. [2]

    A Random Matrix Perspective on Mixtures of Nonlinearities in High Dimensions

    Ben Adlam, Jake A. Levinson, and Jeffrey Pennington. “A Random Matrix Perspective on Mixtures of Nonlinearities in High Dimensions”. In: Proceedings of The 25th International Conference on Artificial Intelligence and Statistics. Ed. by Gustau Camps-Valls, Francisco J. R. Ruiz, and Isabel Valera. V ol. 151. Proceedings of Machine Learning Research. PMLR, 2...

  5. [3]

    Complex analysis

    Lars Valerian Ahlfors and Lars V Ahlfors. Complex analysis. V ol. 3. McGraw-Hill New York, 1979

  6. [4]

    Generalization of two-layer neural networks: An asymptotic viewpoint

    Jimmy Ba et al. “Generalization of two-layer neural networks: An asymptotic viewpoint”. In: International conference on learning representations. 2020

  7. [5]

    High-dimensional analysis of double descent for linear regression with random projections

    Francis Bach. High-dimensional analysis of double descent for linear regression with random projections. 2023. arXiv: 2303.01372 [cs.LG] . URL: https://arxiv.org/abs/2303. 01372

  8. [6]

    Spectral analysis of large dimensional random matrices

    Zhidong Bai and Jack W Silverstein. Spectral analysis of large dimensional random matrices. V ol. 20. Springer, 2010

Show all 70 references
  1. [7]

    Silverstein

    Jinho Baik and Jack W. Silverstein. Eigenvalues of Large Sample Covariance Matrices of Spiked Population Models. 2004. arXiv: math/0408165 [math.ST]. URL: https://arxiv. org/abs/math/0408165

  2. [8]

    Neural networks and principal component analysis: Learning from examples without local minima

    Pierre Baldi and Kurt Hornik. “Neural networks and principal component analysis: Learning from examples without local minima”. In: Neural networks 2.1 (1989), pp. 53–58

  3. [9]

    Benign overfitting in linear regression

    Peter L. Bartlett et al. “Benign overfitting in linear regression”. In:Proceedings of the National Academy of Sciences 117.48 (Apr. 2020), pp. 30063–30070. ISSN : 1091-6490. DOI: 10.1073/ pnas.1907378117. URL: http://dx.doi.org/10.1073/pnas.1907378117

  4. [10]

    Reconciling modern machine-learning practice and the classical bias–variance trade-off

    Mikhail Belkin et al. “Reconciling modern machine-learning practice and the classical bias–variance trade-off”. In: Proceedings of the National Academy of Sciences 116.32 (2019). ISSN : 1091-6490. DOI: 10.1073/pnas.1903070116 . URL: http://dx.doi.org/10. 1073/pnas.1903070116

  5. [12]

    Random Matrix Methods for Machine Learning

    Romain Couillet and Zhenyu Liao. Random Matrix Methods for Machine Learning. https: //zhenyu-liao.github.io/book/ . Cambridge University Press, 2022. DOI: 10.1017/ 9781009128490

  6. [13]

    High-dimensional asymptotics of denoising autoencoders

    Hugo Cui and Lenka Zdeborová. “High-dimensional asymptotics of denoising autoencoders”. In: Advances in Neural Information Processing Systems 36 (2023), pp. 11850–11890

  7. [14]

    A U-turn on Double Descent: Rethinking Parameter Counting in Statistical Learning

    Alicia Curth, Alan Jeffares, and Mihaela van der Schaar. “A U-turn on Double Descent: Rethinking Parameter Counting in Statistical Learning”. In: Thirty-seventh Conference on Neural Information Processing Systems. 2023. URL: https://openreview.net/forum? id=O0Lz8XZT2b

  8. [15]

    Procedures for Reduced-Rank Regression

    P. T. Davies and M. K-S. Tso. “Procedures for Reduced-Rank Regression”. In: Journal of the Royal Statistical Society. Series C (Applied Statistics) 31.3 (1982), pp. 244–255. ISSN : 00359254, 14679876. URL: http : / / www . jstor . org / stable / 2347998(visited on 01/30/2025)

  9. [16]

    On the empirical distribution of eigenvalues of large dimensional information-plus-noise-type matrices

    R Brent Dozier and Jack W Silverstein. “On the empirical distribution of eigenvalues of large dimensional information-plus-noise-type matrices”. In: Journal of Multivariate Analysis 98.4 (2007), pp. 678–694

  10. [17]

    Probability: theory and examples

    Rick Durrett. Probability: theory and examples. V ol. 49. Cambridge university press, 2019. 10

  11. [18]

    The approximation of one matrix by another of lower rank

    Carl Eckart and Gale Young. “The approximation of one matrix by another of lower rank”. In: Psychometrika 1.3 (1936), pp. 211–218

  12. [19]

    The rank of a random matrix

    Xinlong Feng and Zhinan Zhang. “The rank of a random matrix”. In: Applied mathematics and computation 185.1 (2007), pp. 689–694

  13. [20]

    No Double Descent in Prin- cipal Component Regression: A High-Dimensional Analysis

    Daniel Gedon, Antonio H. Ribeiro, and Thomas B. Schön. “No Double Descent in Prin- cipal Component Regression: A High-Dimensional Analysis”. In: Proceedings of the 41st International Conference on Machine Learning. Ed. by Ruslan Salakhutdinov et al. V ol. 235. Proceedings of M...

  14. [21]

    Asymptotic errors for convex penalized linear regression beyond Gaussian matrices

    Cédric Gerbelot, Alia Abbara, and Florent Krzakala. Asymptotic errors for convex penalized linear regression beyond Gaussian matrices. 2020. arXiv: 2002.04372 [stat.ML] . URL: https://arxiv.org/abs/2002.04372

  15. [22]

    Surprises in High-Dimensional Ridgeless Least Squares Interpolation

    Trevor Hastie et al. Surprises in High-Dimensional Ridgeless Least Squares Interpolation

  16. [23]

    Deep Residual Learning for Image Recognition

    Kaiming He et al. Deep Residual Learning for Image Recognition. 2015. arXiv: 1512.03385 [cs.CV]. URL: https://arxiv.org/abs/1512.03385

  17. [24]

    Denoising Diffusion Probabilistic Models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. 2020. arXiv: 2006.11239 [cs.LG]. URL: https://arxiv.org/abs/2006.11239

  18. [25]

    No Double Descent in Self-Supervised Learning

    Dulhan Hansaja Jayalath, Alisia Maria Lupidi, and Yonatan Gideoni. No Double Descent in Self-Supervised Learning . 2023. URL: https : / / openreview . net / forum ? id = qNJRvdKDGYg

  19. [26]

    On the Distribution of the Largest Eigenvalue in Principal Components Analysis

    Iain M. Johnstone. “On the Distribution of the Largest Eigenvalue in Principal Components Analysis”. In: The Annals of Statistics 29.2 (2001), pp. 295–327. ISSN : 00905364, 21688966. URL: http://www.jstor.org/stable/2674106 (visited on 01/26/2025)

  20. [28]

    Deep Learning without Poor Local Minima

    Kenji Kawaguchi. Deep Learning without Poor Local Minima . 2016. arXiv: 1605.07110 [stat.ML]. URL: https://arxiv.org/abs/1605.07110

  21. [29]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. “Learning multiple layers of features from tiny images”. In: (2009)

  22. [30]

    Does Double Descent Occur in Self- Supervised Learning?

    Alisia Lupidi, Yonatan Gideoni, and Dulhan Jayalath. “Does Double Descent Occur in Self- Supervised Learning?” In: arXiv preprint arXiv:2307.07872 (2023)

  23. [31]

    Reduced rank ridge regression and its kernel extensions

    Ashin Mukherjee and Ji Zhu. “Reduced rank ridge regression and its kernel extensions”. In: Statistical analysis and data mining: the ASA data science journal 4.6 (2011), pp. 612–622

  24. [32]

    Deep Double Descent: Where Bigger Models and More Data Hurt

    Preetum Nakkiran et al. Deep Double Descent: Where Bigger Models and More Data Hurt

  25. [33]

    Learning dynamics of linear denoising autoencoders

    Arnu Pretorius, Steve Kroon, and Herman Kamper. “Learning dynamics of linear denoising autoencoders”. In: International Conference on Machine Learning. PMLR. 2018, pp. 4141– 4150

  26. [34]

    Multiple Descents in Unsupervised Learning: The Role of Noise, Domain Shift and Anomalies

    Kobi Rahimi, Tom Tirer, and Ofir Lindenbaum. “Multiple Descents in Unsupervised Learning: The Role of Noise, Domain Shift and Anomalies”. In:arXiv preprint arXiv:2406.11703 (2024)

  27. [35]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. 2015. arXiv: 1505.04597 [cs.CV]. URL: https://arxiv. org/abs/1505.04597

  28. [36]

    Smallest singular value of a random rectangular matrix

    Mark Rudelson and Roman Vershynin. “Smallest singular value of a random rectangular matrix”. In: Communications on Pure and Applied Mathematics: A Journal Issued by the Courant Institute of Mathematical Sciences 62.12 (2009), pp. 1707–1739

  29. [37]

    On the Empirical Distribution of Eigenvalues of a Class of Large Dimensional Random Matrices

    J.W. Silverstein and Z.D. Bai. “On the Empirical Distribution of Eigenvalues of a Class of Large Dimensional Random Matrices”. In: Journal of Multivariate Analysis 54.2 (1995), pp. 175–192. ISSN : 0047-259X. DOI: https://doi.org/10.1006/jmva.1995.1051. URL: https://www.science...

  30. [38]

    Training Data Size Induced Double Descent For Denoising Neural Networks and the Role of Training Noise Level

    Rishi Sonthalia and Raj Rao Nadakuditi. Training Data Size Induced Double Descent For Denoising Neural Networks and the Role of Training Noise Level . 2022. URL: https:// openreview.net/forum?id=5ALGcXpmFyC

  31. [39]

    Topics in random matrix theory

    Terence Tao. Topics in random matrix theory. V ol. 132. American Mathematical Soc., 2012. 11

  32. [40]

    Dimensionality Reduction, Regulariza- tion, and Generalization in Overparameterized Regressions

    Ningyuan Teresa, David W. Hogg, and Soledad Villar. “Dimensionality Reduction, Regulariza- tion, and Generalization in Overparameterized Regressions”. In:SIAM Journal on Mathematics of Data Science 4.1 (Feb. 2022), pp. 126–152. ISSN : 2577-0187. DOI: 10.1137/20m1387821. URL: h...

  33. [41]

    Pure and Spurious Critical Points: a Geomet- ric Study of Linear Networks

    Matthew Trager, Kathlén Kohn, and Joan Bruna. Pure and Spurious Critical Points: a Geomet- ric Study of Linear Networks. 2020. arXiv: 1910.01671 [cs.LG] . URL: https://arxiv. org/abs/1910.01671

  34. [42]

    Why are Big Data Matrices Approximately Low Rank?

    Madeleine Udell and Alex Townsend. Why are Big Data Matrices Approximately Low Rank?

  35. [43]

    High-dimensional probability: An introduction with applications in data science

    Roman Vershynin. High-dimensional probability: An introduction with applications in data science. V ol. 47. Cambridge university press, 2018

  36. [44]

    Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion

    Pascal Vincent et al. “Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion.” In: Journal of machine learning research11.12 (2010)

  37. [45]

    The weighted Moore–Penrose inverse of modified matrices

    Yimin Wei. “The weighted Moore–Penrose inverse of modified matrices”. In:Applied Math- ematics and Computation 122.1 (2001), pp. 1–13. ISSN : 0096-3003. DOI: https://doi. org/10.1016/S0096- 3003(00)00007- 2 . URL: https://www.sciencedirect.com/ science/article/pii/S0096300300000072

  38. [46]

    Optimal exact least squares rank minimization

    Shuo Xiang et al. “Optimal exact least squares rank minimization”. In: Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining. 2012, pp. 480–488

  39. [47]

    Critical Points of Neural Networks: Analytical Forms and Landscape Properties

    Yi Zhou and Yingbin Liang. Critical Points of Neural Networks: Analytical Forms and Landscape Properties. 2017. arXiv: 1710.11205 [stat.ML]. URL: https://arxiv.org/ abs/1710.11205. Appendix We include additional material in the appendix. Section A discusses further related wor...

  40. [51]

    The above eq

    (Empirical Spectral Measure). The above eq. (13) is formally known as Empirical Spectral Measure, and for a symmetric matrix M ∈ Rd×d, it is defined as µM = 1 d dX i=1 δλM i

  41. [52]

    For a real probability measureµ with support supp(µ), the Stieltjes Transform of µ is defined as mµ(α) = Z 1 x − α dµ(x), α ∈ C\supp(µ)

    (Stieltjes Transform). For a real probability measureµ with support supp(µ), the Stieltjes Transform of µ is defined as mµ(α) = Z 1 x − α dµ(x), α ∈ C\supp(µ). 14 Note that the stieltjes transform of the empirical spectral measure is given by, mµM (α) = Z 1 x − α dµM(x) = 1 d ...

  42. [53]

    Let A ∈ Rp×p, and x, y ∈ Rp

    (Sherman-Morrison Formula). Let A ∈ Rp×p, and x, y ∈ Rp. For 1 + yT Ax ̸= 0, we have that (A + xyT )−1 = A−1 − A−1xyT A−1 1+yT Ax

  43. [54]

    For α ∈ C\R+, let ˜Q(α) = ˜A ˜AT − αId −1 = NP i=1 ˜ai˜aT i − αId −1 , where ˜A ∈ Rd×N is a i.i.d real gaussian random matrix, whose entries are sampled from N (0, 1)

    (Concentration of Quadratic Forms, Gaussian case). For α ∈ C\R+, let ˜Q(α) = ˜A ˜AT − αId −1 = NP i=1 ˜ai˜aT i − αId −1 , where ˜A ∈ Rd×N is a i.i.d real gaussian random matrix, whose entries are sampled from N (0, 1). Let ˜Q−j(α) be ˜Q(α) with j-th row and column removed. The...

  44. [55]

    For symmetric and positive semi-definite B, M ∈ Rd×d, and α ∈ C\R+, let M := hP j=1 lixixT i for fixed h, and li ∈ R, xi ∈ Rd

    (Minimal Effect of Finite-Rank Perturbations Inside a Trace, due to [37, Lemma 2.6]). For symmetric and positive semi-definite B, M ∈ Rd×d, and α ∈ C\R+, let M := hP j=1 lixixT i for fixed h, and li ∈ R, xi ∈ Rd. Then, we have that | Tr (B − αI)−1 − (B + M − αI)−1 | ≤ h Im(α) ...

  45. [56]

    For A ∈ Rp×q, B ∈ Rq×p, and λ ∈ R\{0}

    (Weinstein-Aronszajn Identity). For A ∈ Rp×q, B ∈ Rq×p, and λ ∈ R\{0}. Then, we have that det(AB − λIp) = (−λ)p−qdet(BA − λIq)

  46. [57]

    For A ∈ Rp×p, U ∈ Rp×q, V ∈ Rq×p, we have that (A + UVT )−1 = A−1 − A−1U(Iq + VT A−1U)−1VT A−1

    (Woodbury Identity). For A ∈ Rp×p, U ∈ Rp×q, V ∈ Rq×p, we have that (A + UVT )−1 = A−1 − A−1U(Iq + VT A−1U)−1VT A−1. C.2 Getting Subspace Information One useful property of the resolvent and the Stieltjes transform is their connection to the eigenstructure of a random matrix. ...

  47. [58]

    Furthermore, B has rB eigenvalues equal to 1 and the rest equal to 0

    B ∈ R(D−rG)×(D−rG) is defined as B := B1B⊤ 1 , where B1 = ((U⊤ GU ˆW2 )i,j)rG+1≤i≤d,rG+1≤j≤d. Furthermore, B has rB eigenvalues equal to 1 and the rest equal to 0

  48. [59]

    D ∈RrG×rG is a diagonal matrix with its diagonal elements consist of rD number of 1’s and rG − rD number of 0’s

  49. [60]

    rD + rB = r ˆW2 . Proof. We analyze the conditions (17) and (18) further. Consider the eigendecomposition of G, given by G = UGΛGU⊤ G, and the singular value decomposition of ˆW2, given by ˆW2 = U ˆW2 Σ ˆW2 V⊤ ˆW2 . Then, PU⊤ G ˆW2 = U⊤ G ˆW2(U⊤ G ˆW2)† = U⊤ G ˆW2 ˆW† 2UG = U⊤...

  50. [61]

    Let (I, B) a tuple, such that |I| + rB = k, for I ∈ I. Then, for ˆW2 and ˆW1 define critical points for (3) if and only if there exist an invertible matrix C ∈ Rk×k and a tuple (I, B) such that ˆW2 and ˆW1 satisfy that: for |I| < k, ˆW2 = UG,I U2 ˜UB1 C, ˆW1 = C−1 UG,I U2 ˜UB1...

  51. [62]

    Furthermore, assume rZ = n

    For some invertible matrix C ∈ Rrk×rk, global minimizers ˆW∗ 2, ˆW∗ 1 are given by ˆW∗ 2 = UG,kC ˆW∗ 1 = C−1U⊤ G,kYZ⊤( ˜Z ˜Z⊤)−1. Furthermore, assume rZ = n. Then the global minimizer W∗ in the ridgeless case is given uniquely4 by, W∗ = lim λ→0 ˆW∗ 2 ˆW∗ 1 = Pk(Y)Z†

  52. [63]

    For λ >0, we have an unique global minimizer in terms of ˆW∗ = ˆW∗ 2 ˆW∗

  53. [64]

    ⇐" direction of the proof follows from a straightforward calculation, and we therefore omit the details. For the

    On the other hand, at λ = 0, there are multiple global minimizers. But for the both cases, it satisfies that all the critical points other than global minima are saddle points. In other words, all the local minima are global minima and other critical points are saddle points. ...

  54. [65]

    Due to the fact that AI a A†(AI a A†)⊤ = UATaU⊤ A = AI a A†, we have that − Tr(AI a A†(Wsc c )⊤) = Tr(AI a A†(AI a A†)⊤) − Tr(AI a A†H⊤(K1)−1H)− Tr(AI a A†H⊤K−⊤ 1 Z(P⊤P)−⊤D ˜U⊤)

    Distributing −AI a A†: For a notational convenience, we define Ta = 1 if i ∈ I a 0 otherwise. Due to the fact that AI a A†(AI a A†)⊤ = UATaU⊤ A = AI a A†, we have that − Tr(AI a A†(Wsc c )⊤) = Tr(AI a A†(AI a A†)⊤) − Tr(AI a A†H⊤(K1)−1H)− Tr(AI a A†H⊤K−⊤ 1 Z(P⊤P)−⊤D ˜U⊤). Note...

  55. [66]

    The second term has also mean 0 due to Lemma E.9, thus only the variance needs to be bounded for this term

    Distributing −AI a A† ˜UDP†: This is − Tr(AI a A† ˜UDP†(Wsc c )⊤) = Tr(AI a A† ˜UDP†( ˜UDP†)⊤) − Tr(AI a A† ˜UDP†(P†)⊤Z⊤K−⊤ 1 H)− Tr(AI a A† ˜UDP†(P†)⊤Z⊤K−⊤ 1 Z(P⊤P)−⊤D ˜U⊤) Note that some terms do not appear here due to Lemma E.8. The second term has also mean 0 due to Lemma ...

  56. [67]

    The first term is, from Lemma E.7, and Lemma 8 of [27], − Tr(AI a A†H⊤K−1 1 H) = − |I a| n η2 trn c Tr((η2 trnD−2 + Ir)−1) + o |I a| n

    Distributing AI a A†H⊤K−1 1 H: This is Tr(AI a A†H⊤K−1 1 H(Wsc c )⊤) = − Tr(AI a A†H⊤K−1 1 H) + Tr(AI a A†H⊤K−1 1 HH⊤K−1 1 H)+ Tr(AI a A†H⊤K−1 1 HH⊤K−1 1 Z(P⊤P)−1D ˜U⊤) Again, using Lemma E.8, some terms are filtered out. The first term is, from Lemma E.7, and Lemma 8 of [27],...

  57. [68]

    The first term has zero mean, as shown in Lemma E.9

    Distributing AI a A†H⊤K−1 1 ZP†: This is Tr(AI a A†H⊤K−1 1 ZP†(Wsc c )⊤) = − Tr(AI a A†H⊤K−1 1 ZP†(P†)⊤DU⊤)+ Tr(AI a A†H⊤K−1 1 ZP†(P†)⊤Z⊤K−1 1 H)+ Tr(AI a A†H⊤K−1 1 ZP†(P†)⊤Z⊤K−1 1 Z(P⊤P)−1DU⊤) From this point onward, since the proof follows a similar structure to previous arg...

  58. [69]

    For the element-wise variance, Lemma E.3 and Lemmas 4, 6, 7, and 8 from [27] imply that it is of order o(1)

    Distributing AI a A† ˜UD(P⊤P)−1Z⊤K−1 1 H: This is Tr(AI a A† ˜UD(P⊤P)−1Z⊤K−1 1 H(Wsc c )⊤) = − Tr(AI a A† ˜UD(P⊤P)−1Z⊤K−1 1 H)− Tr(AI a A† ˜UD(P⊤P)−1Z⊤K−1 1 HH⊤K−1 1 H)+ Tr(AI a A† ˜UD(P⊤P)−1Z⊤K−1 1 HH⊤K−1 1 Z(P⊤P)−1D ˜U⊤) The first and second terms have zero mean, due to Lemm...

  59. [70]

    Distributing AI a A†UD(P⊤P)−1Z⊤K−1 1 ZP†: This is Tr(AI a A† ˜UD(P⊤P)−1Z⊤K−1 1 ZP†(Wsc c )⊤) = − Tr(AI a A† ˜UD(P⊤P)−1Z⊤K−1 1 ZP†(P†)⊤D ˜U⊤)+ Tr(AI a A† ˜UD(P⊤P)−1Z⊤K−1 1 ZP†(P†)⊤Z⊤K−1 1 H)+ Tr(AI a A† ˜UD(P⊤P)−1Z⊤K−1 1 ZP†(P†)⊤Z⊤K−1 1 Z(P⊤P)−1D ˜U⊤) The second term has zero m...

  60. [2018]

    URL: https://arxiv.org/abs/1705.07474

    arXiv: 1705.07474 [cs.LG]. URL: https://arxiv.org/abs/1705.07474

  61. [2019]

    URL: https://arxiv.org/abs/1912.02292

    arXiv: 1912.02292 [cs.LG]. URL: https://arxiv.org/abs/1912.02292

  62. [2020]

    URL: https://arxiv.org/abs/1903.08560

    arXiv: 1903.08560 [math.ST]. URL: https://arxiv.org/abs/1903.08560

Pith tools

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