acceptodds
Under review as a conference paper at ICLR 2027

Taming Latent Ill-Conditioning in Gaussian Graphical Model Estimation

Abstract

Latent-variable Gaussian graphical models separate a sparse conditional graph from the collective influence of unobserved variables. Although low-rank factorization makes this decomposition convenient to optimize, latent effects of unequal strength can slow gradient descent even when the observed precision matrix is well-conditioned. To address this difficulty, we combine scaled gradient descent with symmetric hard thresholding and a joint feasibility and descent line search. At the correctly specified rank, we establish a local linear convergence bound for the separate sparse and low-rank matrix errors, up to statistical accuracy. The contraction factor and statistical-error multiplier are independent of the latent eigenvalues, with precision conditioning and sparse–low-rank transversality held fixed. The analysis connects an exact cancellation in the scaled gradient to the geometry of fixed-rank matrices, allowing both components to be controlled along the same trajectory. We further provide sufficient conditions under which a thresholded spectral initializer identifies the support and enters the convergence neighborhood. In simulations, ScaledGD maintains similar late-stage convergence rates as latent conditioning worsens while gradient descent slows, and runs faster than AltGD in the sampled-data comparisons.

Primary area:
optimization
Keywords:
Gaussian Graphical Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_mK24OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a scaled gradient descent (ScaledGD) algorithm to estimate latent-variable Gaussian graphical models via a sparse–low-rank decomposition of the precision matrix. The algorithm combines Gram-matrix preconditioning, symmetric hard thresholding, and a joint feasibility-preserving line search. Its main theoretical contribution is a local linear convergence guarantee for both sparse and low-rank components, with a contraction factor independent of the latent condition number under appropriate identifiability and initialization assumptions. The paper also establishes sufficient conditions for spectral initialization and derives finite-sample estimation bounds. Synthetic experiments demonstrate improved convergence under latent ill-conditioning and compare ScaledGD with existing optimization methods.

Critical strengths

Meaningful theoretical contribution. Theorem 1 establishes local linear convergence for the joint sparse and low-rank estimation problem, with a contraction factor independent of the latent condition number. The analysis explicitly handles hard thresholding, support and rank preservation, positive definiteness, and line-search acceptance, extending existing scaled-gradient results to the coupled estimation setting.

Technically well-developed analysis. The fixed-rank geometric arguments in Lemmas 6–9 provide a principled explanation for the condition-number-independent convergence rate. The paper also carefully distinguishes the local convergence guarantee from the signal-strength requirements governing initialization and statistical accuracy.

Experimental support for improved optimization. The conditioning experiment in Section 5.1 demonstrates nearly constant convergence rates for ScaledGD as the latent condition number increases from 5 to 70, while ordinary gradient descent slows substantially. The sampled-data experiments also show lower mean optimization runtimes than AltGD at every sample size tested.

Critical weaknesses

Limited latent-component recovery. Despite the motivation of separating sparse conditional dependencies from latent influences, Table 1 shows relatively poor estimation of the low-rank component. ScaledGD's relative error exceeds 1 at (n=800) and (n=1600), meaning it performs worse on this metric than the trivial zero estimator. The convex PPA baseline consistently achieves substantially lower latent-component error across all evaluated sample sizes. AltGD also exhibits relatively large latent-component errors (1.674, 1.086, 0.775, and 0.549), suggesting that this limitation is not specific to Gram scaling. These results limit the demonstrated effectiveness of the factorized estimation approach for recovering the latent-variable decomposition and warrant further discussion.

Incomplete validation of the central optimization advantage. The condition-number-independence claim is experimentally demonstrated only against ordinary gradient descent using the population covariance and true precision matrix for initialization. The comparison against AltGD, the more directly relevant existing method, is conducted at a single latent condition number κL=5\kappa_L=5. Consequently, the experiments do not establish whether ScaledGD provides a meaningful conditioning-related advantage over AltGD under sampling noise. Moreover, the theoretical guarantee covers equal step sizes, whereas the sampled-data experiments tune the sparse and factor step sizes separately.

Restrictive assumptions and limited practical evaluation. The convergence guarantee requires the correct latent rank, correct initial sparse support, and initialization within a neighborhood whose radius depends on the smallest nonzero latent eigenvalue and minimum sparse-edge strength. The spectral initialization guarantee also relies on thresholds involving unknown population quantities. The sampled-data experiments assume knowledge of the true rank and sparsity budget, and no experiments investigate robustness to their misspecification. All evaluations use synthetic data, leaving practical applicability insufficiently demonstrated.

Questions for authors

The consistently inferior latent-component recovery relative to PPA raises concerns about the factorized approach's practical effectiveness for its motivating decomposition task. Can you explain that? Overall, the theoretical contribution is promising, but the experimental validation and demonstrated practical advantages are not sufficiently compelling to warrant acceptance in the current form.

4 comments

Thank you for the detailed comments. We have revised the paper and added theoretical analysis and experiments.

1. Latent recovery and relative error above 1 in small samples

Even with the true rank, weak latent directions can be difficult to distinguish from sampling noise. Fitting them with r=r⋆r=r^\star can introduce enough noise to push the relative low-rank error above 1. This reflects finite-sample estimation difficulty. Even after optimization converges, the estimate may remain far from the truth. PPA's nuclear-norm regularization shrinks weak spectral directions, helping explain its better small-sample latent recovery.

Data-driven selection supports this explanation. At n=1600,κL=5n=1600,\kappa_L=5, validation selects r=2r=2 instead of r⋆=5r^\star=5, reducing ScaledGD's latent error from 1.117 to 0.649 and precision error from 0.129 to 0.114. See appendix Table 2 and Figure 3 in Section 5.3 for recovery, and Appendix G.3 and Figure 4 for spectral analysis.

2. Condition-number advantage

We extended the sampled-data recovery experiments to larger latent condition numbers. At n=108n=10^8, increasing κL\kappa_L from 5 to 70 raises AltGD's iteration count about 14-fold, compared with 1.7-fold for ScaledGD, which also accurately recovers both components (Section 5.2, Table 1).

Our spectral analysis explains the limited slowdown in the original small-sample experiments. Sampling error masks weak latent directions, so the estimated matrix does not reflect the true ill-conditioning. At n=1600n=1600, the initial condition number stays around 1.4 as the true value increases from 5 to 70. With more samples, the estimated spectrum reflects the true ill-conditioning and AltGD's condition-number dependence becomes apparent (Appendix G.3, Figure 4).

3. Unequal steps and improved convergence bounds

Theorem 1 and its proof in Appendix C now allow unequal positive steps satisfying the step-size bound. We weight the two descent-energy terms separately and define

\tau=\frac{\min\lbrace\eta_S,\eta_B\rbrace}{\eta}.$$ For a fixed step ratio, the contraction factor remains independent of the latent condition number, allowing separate sparse and factor steps as used in the experiments. We also use strong convexity of the log-det loss to directly relate descent energy to the objective gap, tightening the original bound through matrix error. For equal steps at the largest allowed value, the contraction factor improves from $$1-\frac{\gamma^2}{9216\kappa_\ell^2} \quad\text{to}\quad 1-\frac{\gamma}{576\kappa_\ell}.$$ The dependence on the likelihood curvature ratio $\kappa_\ell$ improves from quadratic to linear, while independence from $\kappa_L$ is preserved. See Lemma 9 in Appendix B and Eq. (20) in Appendix A.1. ## 4. Unknown parameters and inaccurate initialization We introduced $\mathrm{ScaledGD}^{\mathrm{sel}}$, which selects rank and sparsity using validation negative log-likelihood. It identifies the true rank and sparsity budget with large samples and improves small-sample recovery by reducing model complexity (Sections 5.2–5.3 and Appendix G.4). We also tested controlled initial support perturbations. With true rank and sparsity fixed, we replace up to 20 true edges and vary $\kappa_L=5,10,\ldots,70$. ScaledGD corrects these support errors and converges across the tested condition numbers. Error curves show similar decay after support recovery, supporting robustness to support mismatch and the condition-number advantage (Section 5.3, Figure 2).
Reviewer_mK24OpenReview ↗

I would regard the rebuttal as successful overall.

Thanks for addressing my comments. The central optimization concern has been convincingly addressed; the equal-step theoretical restriction has been resolved in the revised theorem, and the initialization concern has meaningful additional evidence. The latent-component recovery criticism remains partially valid, but it is no longer sufficient on its own to justify the original weak rejection recommendation. So, I will increase my rating to 3, weak accept.

Thank you for your feedback and for reconsidering our work. We appreciate your suggestions and are glad that our revisions have addressed your main concerns.

Thank you again for your positive feedback. We noticed that the rating in the Official Review still appears unchanged, so we just wanted to check whether the update has gone through. We appreciate your time and consideration.

Reviewer_giBLOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies Gaussian graphical model estimation with a sparse conditional graph and low-rank latent effects. Unequal latent strengths can slow ordinary factor gradient descent even when the observed precision is well-conditioned. The authors rescale the factor updates to speed up progress along weak latent directions. Their main result gives local convergence of both matrix components up to statistical accuracy. The contraction factor is independent of the latent condition number κL\kappa_L, with precision conditioning and sparse-low-rank separation held fixed.

I recommend acceptance. The method is intuitive and a good continuation of previous related work on the same problem. Removing latent-conditioning dependence from the joint local rate is useful, and the experiments support the optimization improvement.

Critical strengths
  1. Theorem 11 removes dependence on the latent condition number from the local convergence rate for recovering both the sparse and low-rank components. This is a useful improvement over ordinary factor gradient descent. Controlling both component errors matters because accurately estimating their difference alone does not ensure their recovery.

  2. The experiments supports the evidence. Figure 11 directly supports the predicted improvement. ScaledGD maintains similar late-stage convergence slopes as κL\kappa_L increases from 55 to 7070, while GD slows substantially. Table 1 also supports a computational benefit. At n=6400n=6400, ScaledGD and AltGD have nearly identical sparse and precision errors, with mean runtimes of 0.24 and 1.31 seconds respectively.

Critical weaknesses
  1. The guarantee requires accurate initialization with the correct support, rank, and sparsity budget. The initialization radius and admissible noise still depend on the weakest latent eigenvalue. The authors acknowledge these restrictions, which limit the scope of the spectral independence.

  2. The conditioning sweep uses the population covariance, while the main sampled-data comparisons fix κL=5\kappa_L=5. The practical evidence under severe latent conditioning and sampling noise is therefore limited. Though as a theoretical contribution the work still stands.

Minor comment that did not affect my score:

Before Lemma 3, briefly recalling uu, vv, and ζ\zeta would help readability. They are defined in Section 3.13.1, but several results intervene.

Questions for authors
  1. Does the acceleration remain visible at larger κL\kappa_L under sampled data?
  2. PPA has the lowest latent-matrix error throughout Table 11. What explains this difference alongside ScaledGD's runtime and sparse-recovery results?
  3. Could you distinguish assumptions needed for identifiability from those used for initialization and local convergence? A short justification, with references to necessity results where available, would help.
4 comments

Thank you for the suggestions. We have revised the paper and added theoretical analysis and experiments.

W1 — Scope of the guarantee

The assumptions on the correct rank, sparsity budget, and initial support are introduced to establish the local convergence guarantee. They ensure that the iterates remain within the intended sparse and fixed-rank structure, allowing us to control the coupled component errors through transversality. These assumptions are not what makes the convergence rate independent of latent conditioning. That property follows from the Gram-scaled descent analysis: the contraction factor has no explicit dependence on rank, sparsity, or the latent eigenvalues.

The initialization radius and admissible noise still depend on the weakest latent eigenvalue, which restricts the scope of the local guarantee but does not affect the spectral independence of the contraction factor within that neighborhood.

We further examined these restrictions experimentally through data-driven rank and sparsity selection and controlled initial support perturbations. ScaledGD achieves effective recovery without oracle model complexity and successfully corrects initial support errors in the tested settings. These results demonstrate robustness beyond the sufficient conditions of our current theory. See Sections 5.2–5.3 and Appendix G.4.

Q1, W2 — Wider condition-number range

We redesigned the convergence and condition number experiment with κL=1,…,140\kappa_L=1,\ldots,140, keeping the precision condition number at 2 and fixing the sparse support and low-rank tangent space. Both methods share initialization, step sizes, and stopping criteria. ScaledGD requires 90 iterations throughout, while ordinary GD increases from 90 to 1567, supporting condition-number-independent convergence (Section 5.1, Figure 1).

The sampled-data recovery experiments cover κL=5,20,50,70\kappa_L=5,20,50,70, with 30 replications per setting and shared sample covariances across methods. At n=108n=10^8, increasing κL\kappa_L from 5 to 70 raises AltGD's mean iteration count about 14-fold, compared with 1.7-fold for ScaledGD (Section 5.2, Table 1).

In small samples, noise masks weak latent directions, so the estimated matrix does not reflect the true ill-conditioning. As the true condition number rises from 5 to 70, the median initial condition number at n=1600n=1600 changes only from 1.36 to 1.38, and AltGD's final condition number remains around 1.44. At n=108n=10^8, the final condition number instead tracks the truth up to about 73, and AltGD's slowdown becomes apparent. This explains the different behavior across sample sizes (Appendix G.3, Figure 4).

Q2 — PPA's latent recovery and precision error

Under limited samples, weak latent directions can be difficult to distinguish from sampling noise. ScaledGD uses a prescribed rank, and even setting it to the true rank may lead to overfitting these weak directions, resulting in large latent recovery errors. In contrast, PPA employs nuclear-norm regularization, which shrinks weak spectral directions and mitigates this effect. This helps explain PPA's better latent recovery in small-sample settings.

We introduced validation-based rank selection for ScaledGD. At (n=1600,κL=5n=1600,\kappa_L=5), validation selects (r=2r=2) instead of the true rank (r⋆=5r^\star=5), reducing the relative latent error from 1.117 to 0.649 and the precision error from 0.129 to 0.114. This supports our explanation that prescribing the true rank can be statistically suboptimal under limited samples, while data-driven model selection improves recovery.

Although PPA achieves a lower latent error in this setting (0.583 versus 1.117 for ScaledGD with the true rank), ScaledGD achieves better sparse recovery (0.115 versus 0.162) and precision recovery (0.129 versus 0.163), alongside its computational advantage. See Appendix Table 2, Section 5.3 (Figure 3), and Appendix G.3.

Q3 — Roles of the assumptions

Appendix A.2 now distinguishes three roles.

  • Identifiability. A trivial intersection of the sparse space and low-rank tangent space makes the linearized decomposition map injective and ensures local identifiability on the fixed-support, fixed-rank model. The transversality margin quantifies stability.
  • Initialization. The edge-strength condition in Theorem 2, Section 4, separates true edges from latent entries and pilot error. The spectral error bound relative to the weakest latent eigenvalue ensures a full-rank factor at the supplied true rank. The remaining accuracy conditions place the start within the local neighborhood.
  • Local convergence. Theorem 1 starts locally with correct support and rank. Precision bounds control likelihood curvature, transversality relates precision and component errors, and step-size and noise bounds preserve positive definiteness and the neighborhood, yielding linear convergence to statistical accuracy.

We also added a reminder of the definitions of uu, vv, and ζ\zeta before Lemma 3, as suggested.

Dear Reviewer giBL,

As the discussion period is coming to an end, we would appreciate your feedback on whether our rebuttal has addressed your concerns. We would be happy to clarify any remaining questions. Thank you for your time and consideration.

Reviewer_giBLOpenReview ↗

Official Comment by Reviewer giBL

Thank you for the response and for adding the sampled-data experiments. The new results address my main question about whether the speedup persists at larger latent condition numbers under sampling noise. The discussion also makes it clearer which parts of the local guarantee still depend on the weakest latent eigenvalue.

I don't have any further questions, and I'll keep my recommendation.

Thank you for your feedback and for taking the time to review our revisions. We appreciate your recognition of the additional results and clarifications.

Reviewer_XAgMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper applies scaled gradient descent (Gram-preconditioned factor updates) to factorized sparse-plus-low-rank LVGGM estimation. It uses symmetric hard thresholding and a joint feasibility/descent line search. At the true rank and after support identification, the authors prove local linear convergence of both component errors, up to statistical error. The contraction factor does not depend on the latent condition number κ_L. They also give sufficient conditions for the spectral initializer and a finite-sample bound. Synthetic experiments compare ScaledGD with GD, AltGD, PPA, and graphical lasso.

Critical strengths
  1. The problem is well chosen and clearly isolated. Example (2) shows that κL\kappa_L can be arbitrarily large while Θ* has condition number 2. This makes clear that the bottleneck is factor conditioning, not precision conditioning.
Critical weaknesses
  1. The algorithm is essentially Xu et al. (2017) AltGD plus hard thresholding, with Tong et al.'s preconditioner added. Under Theorem 1's hypotheses, the support is identified at initialization and hard thresholding keeps exactly the true pairs (Eq. 33). The sparse update therefore reduces to a projected gradient step on a fixed support. As a result, the "joint sparse–low-rank" difficulty in the local regime is milder than the abstract suggests.

  2. At the largest permitted step, ρ=1−γ2/(9216κℓ2)\rho = 1 - \gamma^2 / (9216 \kappa_\ell^2) with κ]ell=16(M/m)2\kappa_]ell = 16(M/m)^2. Even for perfectly conditioned Θ\Theta, this needs on the order of 10610^6 iterations per e-fold. The result is qualitative (independent of κL\kappa_L) but says little about speed. It is not compared quantitatively with AltGD's κL\kappa_L-dependent rate.

  3. All finite-sample comparisons (Table 1, Fig. 2) use κL=5\kappa_L = 5, where preconditioning should matter least. In Table 1, ScaledGD's relative latent error is 1.365 at n = 800 and 1.055 at n = 1600. Both are worse than the trivial estimator L^=0\hat L = 0. PPA wins on L^\hat L at every n, and graphical lasso wins on Θ^\hat \Theta at n = 800. This weakens the claim that both components are recovered in practice.

  4. The rank is known, ss is supplied to the initializer in Sec 5.2, and the initializer's threshold depends on unknown population bounds (ℓ∞,δ∞)(\ell_\infty, \delta_\infty). The authors acknowledge that support and rank selection are left open (Sec 6). However, the paper does not discuss existing LVGGM estimators that avoid exactly this tuning. For example, Yu & Dauwels (MLSP 2015, Variational Bayes learning of graphical models with hidden variables) treat the sparse and low-rank penalty parameters as random variables and learn them by variational Bayes. They report accuracy comparable to stability selection at orders-of-magnitude lower cost. The paper's reported wall-clock advantage assumes oracle (r∗,s∗)(r^*, s^*) for ScaledGD, while PPA must run a validation grid. A fair end-to-end comparison would include the cost of selecting these quantities, and a tuning-free Bayesian LVGGM baseline would be the natural reference point. No real dataset is used, although Xu et al. (2017) used genomic data. Wall-clock comparisons mix implementations (PPA vs. Python ScaledGD), and ScaledGD tunes two step sizes while AltGD tunes one.

  5. Tong, Ma & Chi (2021) already analyze ScaledGD with iteration-varying thresholding for robust PCA, a sparse-plus-low-rank problem. The related-work section should state precisely what changes when the loss is the Gaussian log-likelihood with a PD constraint, compared with a quadratic loss. Further relevant works are not cited:

  • Zhang, Wang & Gu (AISTATS 2018): a unified nonconvex low-rank-plus-sparse framework with double thresholding, under restricted strong convexity and smoothness (RSC/RSS) losses.
  • Xu, Shen, Chi & Ma (ICML 2023), ScaledGD: a damped preconditioner (BTB+λI)−1(B^TB + \lambda I)^{-1} that removes the known-rank requirement while keeping κ\kappa-independent rates. This speaks directly to the rank-misspecification limitation stated in Sec 6.
Questions for authors
  1. Can you add a sampled-data experiment at κL∈{20,50,70}\kappa_L\in \{20, 50, 70\}, with the same Σ^\hat\Sigma shared by ScaledGD and AltGD, that reports iterations and time to reach the statistical floor? This is the regime your theorem targets.

  2. Why does the latent relative error exceed 1 at n ≤ 1600? Is this a scaling-identifiability issue (e.g., mass shifting between the diagonal of S and L) or an initialization effect? Would a diagonal-free sparse space or a different validation criterion change it?

  3. Theorem 1 covers only ηS=ηB\eta_S = \eta_B, yet the experiments tune them separately. Does the proof extend to unequal steps, and does the κL\kappa_L-independence still hold if ηB/ηS\eta_B/\eta_S scales with σ∗\sigma^*?

3 comments

Thank you for the detailed suggestions. We have revised the paper and added theoretical analysis and experiments.

Q1, W2, W4 — Convergence rate and additional experiments

Rate. We tightened the descent-energy bound using strong convexity of the log-det loss. Directly relating descent energy to the objective gap avoids the loss from an intermediate matrix-error bound. For equal steps at the largest allowed value, the contraction factor improves from

\quad\text{to}\quad 1-\frac{\gamma}{576\kappa_\ell}.$$ The dependence on the likelihood curvature ratio $\kappa_\ell$ improves from quadratic to linear, while independence from $\kappa_L$ is preserved. See Theorem 1, Lemma 9 in Appendix B, and Eq. (20) in Appendix A.1. **Larger condition numbers.** We extended the sampled-data experiments to $\kappa_L=5,20,50,70$, sharing the same sample covariance across methods in each replication. At $n=10^8$, increasing $\kappa_L$ from 5 to 70 raises AltGD's mean iteration count about 14-fold, compared with 1.7-fold for ScaledGD, consistent with the predicted condition-number advantage (Section 5.2, Table 1). Our spectral analysis explains the limited slowdown in the original small-sample experiments. Sampling error masks weak latent directions, so the estimated matrix does not reflect the true ill-conditioning. At $n=1600$, the initial condition number stays around 1.4 as the true value increases from 5 to 70. With more samples, the estimated spectrum reflects the true ill-conditioning and AltGD's condition-number dependence becomes apparent (Appendix G.3, Figure 4). **Unknown parameters and computation.** We added $\mathrm{ScaledGD}^{\mathrm{sel}}$, which selects rank and sparsity by validation. It improves small-sample recovery and selects the true rank and sparsity budget at ($n=10^8$). This demonstrates that ScaledGD remains applicable when the true rank and sparsity are unknown. Section 5.3 further examines parameter and initial support perturbations, demonstrating robustness to model misspecification and the ability to correct initial support errors. All methods, including PPA, are implemented in Python. We tune the sparse and factor step sizes separately for both AltGD and ScaledGD. Reported times measure fitting after parameter selection and exclude parameter search. ## Q2, W3 — Relative latent error above 1 in small samples Even with the true rank, weak latent directions can be difficult to distinguish from sampling noise. Fitting them with $r=r^\star$ can introduce enough noise to push the relative low-rank error above 1. This reflects finite-sample estimation difficulty. Even after optimization converges, the estimate may remain far from the truth. At $n=1600,\kappa_L=5$, validation selects $r=2$ instead of $r^\star=5$, reducing the relative low-rank error from 1.117 to 0.649. This supports the explanation above. PPA's nuclear-norm regularization also shrinks weak spectral directions, helping explain its better small-sample latent recovery. See Section 5.3, Figure 3, and appendix Table 2 for recovery results, and Appendix G.3 and Figure 4 for spectral analysis. ## Q3 — Unequal step sizes Theorem 1 and its proof in Appendix C now allow unequal positive steps satisfying the step-size bound. We weight the two descent-energy terms separately and define $$\eta=\max\lbrace\eta_S,\eta_B\rbrace,\quad \tau=\frac{\min\lbrace\eta_S,\eta_B\rbrace}{\eta}.$$ At the largest allowed step, $\rho=1-\tau\gamma/(576\kappa_\ell)$. The rate remains independent of latent eigenvalues for a fixed step ratio. ## W1, W5 — Contribution beyond prior work We establish simultaneous recovery of the coupled sparse and low-rank components with a contraction factor independent of the latent condition number. Existing low-rank ScaledGD theory does not directly give this result. The likelihood depends on both components through $\Theta=S-BB^\top$, so loss decrease alone cannot guarantee recovery of each component. Their errors can cancel even on the correct support. Gram scaling must also work jointly with the sparse update and positive-definiteness constraint. We use transversality to control this cancellation, relate joint descent energy to the objective gap, and prove that the scaled joint updates preserve positive definiteness and achieve local linear convergence. These component recovery guarantees distinguish our analysis from existing AltGD and low-rank ScaledGD results. The two suggested references and discussion appear in the introduction, page 2, lines 82–85, and page 3, lines 108–111.

Thank you for the reply. The unequal-step extension, tightened contraction factor, added citations, and unified Python implementations address some of my concerns. I'll keep my score.

Thank you for your feedback. We appreciate your recognition of the revisions and your helpful suggestions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.