acceptodds
Under review as a conference paper at ICLR 2027

Geometry Determines Local Acceleration in Griffin-Lim Phase Retrieval

Abstract

Acceleration in Griffin–Lim phase retrieval is commonly tuned heuristically, despite substantial instance-to-instance variation in local geometry. We show that this geometry determines the local acceleration problem. At a regular feasible solution, the transverse Griffin–Lim Jacobian is self-adjoint and positive semidefinite, with eigenvalues equal to the squared cosines of the principal angles between the signal-consistency subspace and the tangent space of the magnitude manifold. This geometry–spectrum correspondence turns local acceleration into an explicit spectral design problem. For the three-parameter AGLA family, we derive the exact modal dynamics of the linearization and solve the resulting minimax spectral-radius problem in closed form. The geometry-optimal design achieves a strictly smaller optimal local factor than FGLA, and hence GLA, on every nondegenerate transverse spectral interval. When the local spectrum is unknown, however, acceleration becomes an information problem. We show that passive trajectory observations cannot identify spectral directions they do not excite. Under explicit local active-subspace and consistency conditions, trajectory-based Ritz estimates nevertheless recover the excited dominant spectral edge, yielding an online method that conditionally attains the corresponding conservative geometry-designed asymptotic rate. Real-audio STFT experiments support the predicted geometry–spectrum relationship and reveal a complementary robustness–rate separation: online adaptation does not improve paired finite-window local contraction rates on average, but substantially improves empirical convergence success under the frozen initialization protocol. Across three real-audio evaluations, Online attains 73.30–74.94% success, versus 54.87–55.04% for oracle-tuned FGLA. On a disjoint 4,752-run MUSAN evaluation, the same frozen method attains 73.30%, compared with 52.06% for a literature-configured NAGLA and 47.62% for a published AGLA configuration; a larger 11,856-run MUSAN evaluation reproduces the robustness–rate separation. Together, these results identify local geometry as the organizing principle for Griffin–Lim acceleration, characterize the spectral information available to passive adaptation, and separate geometry-optimal local acceleration from what can be realized online.

Primary area:
optimization
Keywords:
Griffin-Lim algorithm, phase retrieval, spectral acceleration, local convergence, principal angles, adaptive momentum, Ritz estimation, STFT phase retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_SjNCOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

I'm unable to summarize the contributions of this manuscript due to its lack of clarity.

Critical strengths

None

Critical weaknesses

The paper is unreadable to me. As illustrated by the abstract, it reads as an inventory of technical claims whose importance is poorly explained. I don't see any motivation for the problem: "Given the local geometry of a regular solution, what is the best achievable local contraction, and which inertial parameters attain it?" I don't understand what "local geometry" refers to, what an inertial parameter is, and why this technical question matters, what are the insight provided by this study etc. The mathematical exposition doesn't look better; the algorithm is not described in the Problem Setup section; notations such as d2d^2 or PMsP_{M_s} are not defined, etc.

Questions for authors

What audience is this paper written for?

5 comments
AuthorsOpenReview ↗

Weakness 1 and Question 1 - Motivation, Significance and Intended Audience

Weakness 1

The paper is unreadable to me. As illustrated by the abstract, it reads as an inventory of technical claims whose importance is poorly explained. I don't see any motivation for the problem: "Given the local geometry of a regular solution, what is the best achievable local contraction, and which inertial parameters attain it?" I don't understand what "local geometry" refers to, what an inertial parameter is, and why this technical question matters, what are the insight provided by this study etc.

Question 1

What audience is this paper written for?

We thank Reviewer SjNC for this direct assessment. We have rewritten the abstract, introduction and problem setup to explain the motivation before the technical results.

Motivation. In audio phase retrieval, only STFT magnitudes are observed and the phase must be reconstructed. The Griffin–Lim algorithm (GLA) alternates between two constraints: consistency with an underlying signal (projection onto the subspace UU) and agreement with the measured magnitudes (projection onto the set Ms\mathcal M_s). It is simple and widely used, but it converges slowly when the two constraint sets are nearly tangent, because each projection then makes only a small correction. Accelerated variants (FGLA, AGLA) add momentum and relaxation, but their parameters are usually tuned by hand, although the best choice depends on the problem instance.

What "local geometry" and "inertial parameters" mean.

  • Local geometry: the principal angles between the subspace UU and the tangent space VV of Ms\mathcal M_s at a solution. The eigenvalues of the linearized Griffin–Lim map are the squared cosines of these angles (Theorem 3.1). Excluding the trivial global-phase direction, the largest one, L=cos⁡2θ1L=\cos^2\theta_1, determines the slowest local convergence; LL close to 1 means near-tangency.
  • Inertial parameters: the coefficients (α,β,γ)(\alpha,\beta,\gamma) in Section 2, eq. (2.2), which extrapolate along the previous step (α\alpha, β\beta) and relax the update (γ\gamma).

Why the question matters and what we show.

  1. Exact optimal design. The local spectrum is exactly the interval [0,L][0,L], because the radial direction is always a zero mode. On this interval we derive in closed form the unique parameters that minimize the worst-case local rate within AGLA, and prove they are strictly better than optimally tuned FGLA (Theorem 4.1). This replaces hand tuning by a formula in one geometric quantity LL.
  2. What is needed in practice. LL is unknown, so it must be estimated from the iterates. Theorem 4.2 states exactly how accurate the estimate must be (accuracy relative to the gap 1−L1-L), Theorem 5.1 explains when trajectory-based estimates can reach it, and Theorem 6.1 proves local convergence of a safeguarded online method even when estimates are inaccurate.
  3. Experiments. On real-audio STFT problems, the oracle design shows the predicted local acceleration, while the online method mainly improves robustness at matched call budgets but not at matched wall-clock time (Section 7).

Audience. The paper is written for researchers in optimization and iterative inverse problems, including signal and audio processing, who use alternating-projection or momentum methods and want principled parameter choices.

AuthorsOpenReview ↗

Weakness 2 - Mathematical Exposition, Algorithm Description and Notation

Weakness 2

The mathematical exposition doesn't look better; the algorithm is not described in the Problem Setup section; notations such as d2d^2 or PMsP_{\mathcal M_s} are not defined, etc.

We thank the reviewer. Section 2 now states the full setup and algorithm before any theory.

Definitions (Section 2). Let T:Cd→CMT:\mathbb{C}^d\to\mathbb{C}^M be a full-column-rank analysis operator and s∈R+Ms\in\mathbb R_+^M the measured magnitudes. The two constraint sets are U=range(T),Ms={c∈CM:∣ci∣=si for all i}.U=\mathrm{range}(T),\qquad \mathcal M_s=\lbrace c\in\mathbb{C}^M : \lvert c_i\rvert = s_i \text{ for all } i\rbrace . The consistency projection is PU=TT†P_U = TT^\dagger, and the magnitude projection replaces each entry ci≠0c_i \neq 0 of cc by sici/∣ci∣s_i c_i/\lvert c_i\rvert. One Griffin–Lim step is G=PUPMsG = P_U P_{\mathcal M_s}, and the objective is f(t)=12∥∣t∣−s∥2=12dist(t,Ms)2,t∈U(eq. 2.1).f(t)=\tfrac12 \big\lVert \lvert t\rvert - s \big\rVert^2 = \tfrac12 \mathrm{dist}(t,\mathcal M_s)^2,\qquad t\in U \quad \text{(eq. 2.1)}. Convergence is measured against the solution orbit O⋆={eiϕc⋆:ϕ∈R}\mathcal O_\star=\lbrace e^{i\phi}c^\star : \phi\in\mathbb{R}\rbrace, since a global phase cannot be recovered.

Algorithm (eq. 2.2). AGLA keeps the current and previous iterates (t,t−)(t, t^-). With Δ=t−t−\Delta = t - t^-, one step with parameters (α,β,γ)(\alpha,\beta,\gamma) is c=t+αΔ,dβ=t+βΔ,t+=(1−γ)dβ+γG(c).c = t + \alpha\Delta,\qquad d_\beta = t + \beta\Delta,\qquad t^+ = (1-\gamma) d_\beta + \gamma G(c). GLA corresponds to (0,0,1)(0,0,1) and FGLA sets γ=1\gamma = 1. The former notation d2d^2 has been removed; distances are written explicitly as dist(⋅,⋅)\mathrm{dist}(\cdot,\cdot), and dβd_\beta denotes the second extrapolate.

Organization. Problem setup and algorithm (Section 2) → geometry and spectrum (Section 3) → exact acceleration and spectral tolerance (Section 4) → online estimation (Section 5) → safeguarded convergence (Section 6) → experiments (Section 7) → discussion and limitations (Section 8). Each theorem is followed by a short interpretation in the main text; proofs are in Appendices A–G and the experimental protocol is in Appendix H. Figure 1 gives an overview of the method.

We sincerely thank you for your thoughtful and constructive feedback, which has helped us substantially strengthen the manuscript. We believe that the new theoretical results, additional experimental evidence, and clarifications directly address your concerns and further establish the significance of our contributions.

We would greatly appreciate your reassessment of the submission in light of these developments. We welcome any remaining questions or suggestions and look forward to continuing our discussion during the rebuttal period.

Thank you again for your time, expertise, and careful consideration of our work.

Reviewer_k1LbOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents an analysis of local convergence and acceleration of Griffin-Lim through the geometry and spectrum of the linearized map at a feasible solution. The authors derive optimal parameters within the AGLA family, proposes an online spectral estimator, and evaluates local contraction and convergence success on audio phase-retrieval problems.

Critical strengths

The presented connection between local geometry and acceleration is interesting.

Critical weaknesses

It is clear from the definitions that c⋆∈N⋆c^\star\in N_\star, since c⋆∈Uc^\star\in U and c⋆c^\star is orthogonal to ic⋆ic^\star under the real inner product. Moreover, from the geometric derivation of (V), the vector c⋆c^\star is orthogonal to the tangent space of the magnitude manifold, so PVc⋆=0P_Vc^\star=0. Consequently, Theorem 3.1 gives J⋆c⋆=PUPVc⋆=0J_\star c^\star=P_UP_Vc^\star=0, which implies λ−=0\lambda_-=0. If this is correct, I do not see how the one-edge and two-edge oracle factors can differ (lines 353-355), or how the spectral width can shrink at fixed λ+\lambda_+ (lines 261-266). Since these distinctions appear central to the paper's contributions, I believe this would require a substantial restructuring of the paper and a reassessment of its main claims.

I also found the paper very difficult to read. The presentation is verbose and repetitive, making it difficult to identify the main arguments and their significance. In particular, most acronyms are not explained, the related works section is very short and gives practically no insight or intiution on what has previously been done.

Some experimental details are also unclear, particularly the initialization settings, contraction-fitting windows, and convergence criteria.

Questions for authors
  1. Am I overlooking something in the argument that λ−=0\lambda_-=0? If this holds, how does it affect the main contributions of the paper?
6 comments
AuthorsOpenReview ↗

Weakness 1 and Question 1 - The Radial Zero Mode (λ₋ = 0) and Its Effect on the Main Contributions

Weakness 1

It is clear from the definitions that c⋆∈N⋆c^\star \in N_\star, since c⋆∈Uc^\star \in U and c⋆c^\star is orthogonal to ic⋆ic^\star under the real inner product. Moreover, from the geometric derivation of VV, the vector c⋆c^\star is orthogonal to the tangent space of the magnitude manifold, so PVc⋆=0P_V c^\star = 0. Consequently, Theorem 3.1 gives J⋆c⋆=PUPVc⋆=0J_\star c^\star = P_U P_V c^\star = 0, which implies λ−=0\lambda_- = 0. If this is correct, I do not see how the one-edge and two-edge oracle factors can differ (lines 353–355), or how the spectral width can shrink at fixed λ+\lambda_+ (lines 261–266). Since these distinctions appear central to the paper's contributions, I believe this would require a substantial restructuring of the paper and a reassessment of its main claims.

Question 1

Am I overlooking something in the argument that λ−=0\lambda_- = 0? If this holds, how does it affect the main contributions of the paper?

We thank Reviewer k1Lb for this precise observation. The argument is correct. Since c⋆∈U∩(ic⋆)⊥Rc^\star \in U \cap (ic^\star)^{\perp_{\mathbb{R}}} and c⋆⊥RVc^\star \perp_{\mathbb{R}} V, we have PVc⋆=0P_V c^\star = 0 and hence J⋆c⋆=PUPVc⋆=0J_\star c^\star = P_U P_V c^\star = 0. Thus λ−≡0\lambda_- \equiv 0 is a structural property of Griffin–Lim. We withdraw the interpretation in the original submission that the lower edge and the spectral width can vary independently (the passages at lines 261–266 and 353–355), and we restructured the paper around this fact.

Effect on the contributions.

  1. Exact spectrum (Theorem 3.1). The revised theorem states ker⁡J⋆=Rc⋆\ker J_\star = \mathbb{R}c^\star and gives the transverse endpoints λ−=0\lambda_- = 0 and λ+=L:=λmax⁡(J⋆∣N⋆)∈[1/2,1)\lambda_+ = L := \lambda_{\max}(J_\star\vert_{N_\star}) \in [1/2, 1). The one-edge design is therefore the exact full-interval design, not a conservative approximation.
  2. Exact minimax design (Theorem 4.1). On [0,L][0,L], the unique stationary minimax design within the original three-parameter AGLA family is q0(L)=3−L−(1−L)(9−L)2,(α,β,γ)=(q03−q0, 0, 1+q0).q_0(L)=\frac{3-L-\sqrt{(1-L)(9-L)}}{2},\qquad (\alpha,\beta,\gamma)=\Big(\frac{q_0}{3-q_0},\ 0,\ 1+q_0\Big). At the radial endpoint the characteristic polynomial is r(r+q0)r(r+q_0), and at the dominant edge it is (r−q0)2(r-q_0)^2. The Jury conditions give uniqueness, and q0(L)<qF(L)=1−1−L<Lq_0(L) < q_F(L) = 1-\sqrt{1-L} < L, i.e. a strictly smaller worst-case local root factor than optimal FGLA. The radial zero mode is what makes this design exact: it constrains the admissible relaxation.
  3. Spectral tolerance and safeguarding (Theorems 4.2 and 6.1). Theorem 4.2 shows that every frozen estimate ℓ∈[0,1)\ell\in[0,1) gives a locally convergent design, and gives the exact interval of estimates that strictly improve on optimal FGLA. Theorem 6.1 proves local convergence of the checkpointed Online policy under exact sufficient decrease, without requiring estimator consistency.
  4. General two-edge result. The two-edge design with λ−>0\lambda_- > 0 is kept only in Appendix G, as a result for other smooth fixed-point maps with self-adjoint PSD derivatives, checked on a controlled non-STFT example (Table 3).

Numerical checks (Section 7.1). On 24 real-STFT geometries, the maximum radial residual ∥J⋆c⋆∥\lVert J_\star c^\star\rVert is 1.57×10−161.57\times10^{-16}, and independent Jacobian and principal-sine Gram computations agree on the gap to 3.08×10−73.08\times10^{-7} (CPU replication: 3.53×10−73.53\times10^{-7}). For dominant-mode perturbations of size 10−610^{-6}, one-edge AGLA has smaller orbit error than optimal FGLA on all 24 geometries at each of nine checkpoints (median relative error 0.3588 vs. 0.5410 at 360 calls; 0.00211 vs. 0.0170 at 1,440 calls). For radial perturbations, GLA removes the error in one step while optimal AGLA retains the factor −q0-q_0, as predicted by the zero mode.

AuthorsOpenReview ↗

Weakness 2 - Readability, Related Work and Experimental Details

Weakness 2

I also found the paper very difficult to read. The presentation is verbose and repetitive, making it difficult to identify the main arguments and their significance. In particular, most acronyms are not explained, the related works section is very short and gives practically no insight or intuition on what has previously been done. Some experimental details are also unclear, particularly the initialization settings, contraction-fitting windows, and convergence criteria.

We thank the reviewer and have revised the manuscript as follows.

Structure. Sections 1–8 now follow one line of argument: problem setup and full algorithm (Section 2), geometry and spectrum (Section 3), exact acceleration and spectral tolerance (Section 4), online estimation (Section 5), safeguarded convergence (Section 6), experiments (Section 7), and discussion and limitations (Section 8). Figure 1 gives an overview of the oracle design and the Online realization; Figure 2 shows the spectral diagnostics and modal dynamics.

Acronyms and notation. GLA, FGLA, AGLA, NAGLA and RestartNAG are defined at first use, and all operators (PUP_U, PMsP_{\mathcal M_s}, GG), sets (UU, Ms\mathcal M_s, O⋆\mathcal O_\star) and state variables are defined in Section 2.

Related work. The related-work discussion (Section 1) now covers classical alternating projections and Griffin–Lim, local convergence via tangent-space angles, spectral structure in phase retrieval, FGLA/AGLA and Nesterov-type acceleration, polynomial acceleration near fixed-point manifolds, Rayleigh–Ritz compression, sufficient-decrease convergence principles, and STFT injectivity and stability, and states what our analysis does and does not claim relative to these works.

Experimental details (Section 7.2, Appendix H).

  • Setting: 12 MUSAN files (four each of music, noise and speech), 4,096-sample signals, two full-complex periodic STFT profiles: Gaussian (256/32) and Hann (512/128).
  • Initializations: for each file and profile, transverse perturbations of scale 0.001, 0.01, 0.05; phase jitter of scale 0.05, 0.15, 0.30; signal noise of scale 0.1; and random phase. This gives 144 historical, 24 signal-noise and 24 random-phase starts per method, with identical initial states across methods.
  • Convergence criteria: joint success requires e(t)≤0.1e(t0)e(t) \le 0.1 e(t_0) for the normalized orbit error and rmag(t)≤10−5r_{\mathrm{mag}}(t) \le 10^{-5} for the normalized magnitude residual, evaluated at the committed iterate. We also report both criteria separately.
  • Contraction-fitting windows: we no longer fit contraction windows. Local-rate experiments compare measured orbit errors with exact analytical predictions at fixed call counts, and adaptive experiments compare success at matched call and wall-clock budgets.

We sincerely thank the reviewer again for the thoughtful and constructive feedback, which has helped us strengthen the manuscript. We hope that the new theoretical results, numerical evidence, and clarifications address your concerns. We would greatly appreciate your reconsideration of the revised manuscript and welcome any further feedback at your earliest convenience during the discussion period.

Thank you again for your time and careful consideration.

Reviewer_2zMfOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates the local convergence acceleration of the Griffin–Lim phase retrieval algorithm, aiming to address the limitations of existing acceleration methods that rely on heuristically selected parameters and fail to account for instance-specific local geometry. The authors first analyze the geometric relationship between the signal-consistency subspace and the magnitude-constraint manifold, establishing a connection between the principal angles of these spaces and the eigenvalues of the local Jacobian. Based on this relationship, they derive closed-form optimal acceleration parameters and convergence factors for the three-parameter AGLA algorithm, theoretically achieving faster local convergence than conventional GLA and FGLA. To address the unknown local spectral information in practice, the authors further propose an online adaptive method that estimates spectral properties from iteration trajectories using Ritz approximations and adjusts acceleration parameters accordingly. Finally, experiments on real-audio STFT phase retrieval validate the theoretical geometry–spectrum relationship and demonstrate that the proposed online method substantially improves convergence success rates, although it does not achieve a clear advantage in local convergence speed among successfully converged cases.

Critical strengths

Solid theoretical analysis. The paper provides a rigorous mathematical analysis connecting the local geometry of the signal-consistency subspace and magnitude manifold to the convergence behavior of Griffin–Lim algorithms. The derived geometry–spectrum relationship offers a clear theoretical foundation for understanding local acceleration.

Principled acceleration parameter design. Rather than relying on heuristic parameter tuning, the authors derive closed-form optimal parameters for the three-parameter AGLA family and establish theoretical improvements in local convergence factors over GLA and FGLA. This provides a more systematic approach to acceleration design.

Practical extension through online adaptation. The proposed trajectory-based Ritz estimation method addresses the difficulty of obtaining local spectral information in practice, allowing acceleration parameters to be adjusted during iterations without explicitly computing the full Jacobian.

Extensive experimental evaluation. The paper evaluates the proposed approach on real-audio STFT phase retrieval across multiple experimental populations, including comparisons with GLA, FGLA, NAGLA, and AGLA. The results demonstrate substantial improvements in empirical convergence success rates, supported by additional robustness analyses and ablation studies.

Critical weaknesses

Limited novelty and generalizability. The proposed approach builds upon existing alternating-projection theory, principal-angle geometry, and Ritz spectral estimation techniques. Although the theoretical derivations are rigorous, the methodological novelty is relatively limited, and the analysis is primarily restricted to the Griffin–Lim framework without demonstrating broader applicability.

Gap between theoretical analysis and practical improvements. The theoretical guarantees rely on restrictive local regularity and spectral estimation assumptions. More importantly, the experiments show that the online method primarily improves convergence success rates rather than local convergence speed. The connection between the theoretical acceleration results and the observed robustness improvements remains insufficiently explained.

Incomplete experimental evaluation and baseline comparisons. Although multiple baselines are considered, NAGLA and AGLA are evaluated using published parameter configurations without task-specific tuning, potentially limiting the fairness of the comparisons. Furthermore, the paper lacks comprehensive computational efficiency evaluations, such as runtime overhead and convergence speed under equal computational budgets. The experiments are also largely restricted to a single audio dataset.

Poor readability, unclear organization, and insufficient visualization. The manuscript suffers from poor readability and a lack of clear organization. The main text is heavily dominated by mathematical equations and theoretical derivations, with insufficient intuitive explanations, methodological motivation, and textual descriptions to help readers understand the underlying ideas. The connections between different sections and the overall methodological framework are not clearly presented, making the paper difficult to follow. Furthermore, the paper lacks a methodological overview figure or algorithmic flowchart to illustrate the proposed approach, and the experimental visualization is insufficient, with most results concentrated in a single figure and several important comparisons and ablation studies relegated to the appendix. In particular, Figure 1(a) would benefit from a clearer explanation of the discrepancies between geometric predictions and directly computed spectral values. These issues significantly reduce the clarity, accessibility, and overall presentation quality of the paper.

Questions for authors
  1. The proposed method theoretically achieves faster local convergence, but the experiments show no clear improvement in local convergence rates. What evidence supports the claimed practical acceleration?
  2. The theoretical analysis focuses on local convergence, while the main experimental improvement comes from higher convergence success rates. Is there sufficient evidence to establish a connection between the theoretical results and the observed improvements?
  3. Why are NAGLA and AGLA evaluated only with their published parameter configurations? Could the authors provide comparisons with properly tuned baselines to ensure a fair evaluation?
  4. The ablation results show a substantial performance gap between the fixed one-edge oracle and the adaptive method. Are additional experiments needed to isolate the actual contributions of spectral estimation and adaptive parameter updates?
  5. Given the limited experimental settings, could the authors provide additional evaluations on other datasets and report computational costs to better support the generalizability and practical efficiency of the proposed method?
  6. One of the paper's main theoretical contributions is the derivation of optimal acceleration parameters based on the complete spectral interval ([λ−\lambda_-, λ+\lambda_+]). However, the proposed online algorithm only estimates the dominant spectral edge and adopts a conservative interval ([0, λ^+\hat{\lambda}_+]). Consequently, the most distinctive theoretical result is not fully utilized in the practical implementation, raising concerns about its actual contribution to the observed performance improvements.
7 comments
AuthorsOpenReview ↗

Weakness 1 and Question 6 - Novelty, Generalizability and Use of the Complete Spectral Interval

Weakness 1

Limited novelty and generalizability. The proposed approach builds upon existing alternating-projection theory, principal-angle geometry, and Ritz spectral estimation techniques. Although the theoretical derivations are rigorous, the methodological novelty is relatively limited, and the analysis is primarily restricted to the Griffin–Lim framework without demonstrating broader applicability.

Question 6

One of the paper's main theoretical contributions is the derivation of optimal acceleration parameters based on the complete spectral interval [λ−,λ+][\lambda_-,\lambda_+]. However, the proposed online algorithm only estimates the dominant spectral edge and adopts a conservative interval [0,λ^+][0,\hat\lambda_+]. Consequently, the most distinctive theoretical result is not fully utilized in the practical implementation, raising concerns about its actual contribution to the observed performance improvements.

We thank Reviewer 2zMf for these questions, which led to an important clarification in the revision.

Scope of the contribution (Weakness 1). We agree that alternating projections, principal angles and Rayleigh–Ritz compression are classical tools, and we cite them as such (Section 1, Related work). Our contribution is to connect them into an exact design for Griffin–Lim acceleration:

  • Theorem 3.1: at a regular solution, the Griffin–Lim Jacobian J⋆=PUPV∣UJ_\star = P_U P_V\vert_U is self-adjoint PSD, its eigenvalues are the squared cosines of the principal angles between UU and the tangent space VV, and the radial direction is an unavoidable zero mode (J⋆c⋆=0J_\star c^\star = 0). The realizable transverse interval is therefore exactly [0,L][0,L] with L=λmax⁡(J⋆∣N⋆)L=\lambda_{\max}(J_\star\vert_{N_\star}).
  • Theorem 4.1: on [0,L][0,L], the unique stationary minimax design within the three-parameter AGLA family is

q0(L)=3−L−(1−L)(9−L)2,(α,β,γ)=(q03−q0, 0, 1+q0),q_0(L)=\frac{3-L-\sqrt{(1-L)(9-L)}}{2},\qquad (\alpha,\beta,\gamma)=\Big(\frac{q_0}{3-q_0},\ 0,\ 1+q_0\Big),

with q0(L)<qF(L)=1−1−L<Lq_0(L) < q_F(L) = 1-\sqrt{1-L} < L, i.e. a strictly smaller worst-case local root factor than optimally tuned FGLA, and hence GLA. We do not claim optimality over all accelerated methods; the result is exact within this family.

Question 6 (one-edge vs. full interval). Following Reviewer k1Lb's observation that λ−≡0\lambda_- \equiv 0, the interval [0,L][0,L] is not a conservative simplification: it is the exact realizable interval of Griffin–Lim. The one-edge design is therefore the full-interval design, and Online uses the main theoretical result directly; the only unknown is LL. The revision makes this precise:

  • Theorem 4.2 shows that every frozen estimate ℓ∈[0,1)\ell\in[0,1) gives a locally convergent design, and gives the exact interval of estimates that strictly improve on optimal FGLA. As the gap g=1−L→0g=1-L\to 0, the bounds on the estimated-to-true gap ratio approach 1/21/2 and 9/89/8, so accuracy must be measured relative to the spectral gap.
  • Theorem 5.1 bounds the Ritz estimation error in terms of missing dominant-mode excitation, phase leakage and conditioned secant error, i.e. when passive observations can reach this accuracy.
  • A mismatch sweep (576 trajectories, 5,184 checkpoints, with CPU replication) matches the predicted two-step errors (median relative discrepancy 7.0×10−97.0\times10^{-9} for dominant modes; Section 7.1, Figure 2(b), Appendix H).

Generalizability (Weakness 1). Appendix G proves the two-edge minimax design for smooth fixed-point maps with self-adjoint PSD derivatives, where a positive lower edge can occur, and Table 3 checks it on a controlled non-STFT example over four intervals. Our experiments remain focused on STFT phase retrieval, and we state this scope in Section 8.

AuthorsOpenReview ↗

Weakness 2 and Question 1 - Gap Between Theory and Practice, and Evidence of Acceleration

Weakness 2

Gap between theoretical analysis and practical improvements. The theoretical guarantees rely on restrictive local regularity and spectral estimation assumptions. More importantly, the experiments show that the online method primarily improves convergence success rates rather than local convergence speed. The connection between the theoretical acceleration results and the observed robustness improvements remains insufficiently explained.

Question 1

The proposed method theoretically achieves faster local convergence, but the experiments show no clear improvement in local convergence rates. What evidence supports the claimed practical acceleration?

We thank the reviewer for this central question. We agree with the observation, and the revision now separates two claims explicitly: (i) the oracle design accelerates local convergence as predicted, and (ii) Online's practical gains are mainly in robustness (joint success), not in realized local rate.

Assumptions (Theorem 6.1). We no longer rely on estimator consistency. Theorem 6.1 proves local convergence of the checkpointed Online policy under exact sufficient decrease, without requiring estimator consistency or eventual elimination of fallback. Appendix E gives an explicit invariant neighborhood and a finite-budget success certificate after entry. The result is local: it does not establish global convergence or basin enlargement. Appendix E also shows that a fixed positive acceptance threshold can reject an optimal oracle block on near-degenerate trajectories.

Controlled evidence of oracle local acceleration (Section 7.1, Figure 2(c)). Using the numerical spectra, zero initial momentum and dominant-mode perturbations of size 10−610^{-6} on 24 real-STFT geometries, one-edge AGLA attains smaller orbit error than optimal FGLA on all 24 geometries at each of nine checkpoints. Median relative orbit error e(tn)/e(t0)e(t_n)/e(t_0):

Method360 calls1,440 calls
GLA0.99330.9733
Optimal FGLA0.54100.0170
One-edge AGLA0.35880.00211

At 1,440 calls the median error is about 8×8\times smaller than optimal FGLA. Up to 360 calls the maximum relative two-step prediction error is 3.10×10−63.10\times10^{-6} (504 checks). On radial perturbations (Figure 2(d)) the ordering reverses: GLA removes a small radial perturbation in one step, while optimal AGLA retains the factor −q0-q_0. This is consistent with minimax optimality, which is a worst-case and not a pointwise statement. Across four perturbation scales, AGLA outperforms optimal FGLA on 95/96 dominant-mode pairs at 360 calls (Appendix H).

What this does and does not show. This test uses the true numerical edge LL, so it validates the oracle rate of Theorem 4.1, not acceleration by Online. For Online, the mechanism audit (Section 7.3, Table 2) shows that only 1.40% of early synthesized edges (within 360 calls) satisfy the tolerance of Theorem 4.2, and 50.84% of call-mode blocks use the ceiling estimate. Online therefore rarely realizes the oracle rate at practical budgets; its gains come mainly from adaptation and rollback (see our reply to Questions 2 and 4). We state this separation in the abstract and in Section 8.

AuthorsOpenReview ↗

Question 2 and Question 4 - What Explains Robustness, and Which Component Matters

Question 2

The theoretical analysis focuses on local convergence, while the main experimental improvement comes from higher convergence success rates. Is there sufficient evidence to establish a connection between the theoretical results and the observed improvements?

Question 4

The ablation results show a substantial performance gap between the fixed one-edge oracle and the adaptive method. Are additional experiments needed to isolate the actual contributions of spectral estimation and adaptive parameter updates?

We thank the reviewer. The revision adds component ablations and a block-level mechanism audit (Section 7.3; Table 6 in Appendix H).

Component ablation. Joint successes; historical starts n=144n=144, signal-noise starts n=24n=24:

MethodHistorical, 360 callsNoisy, 360 callsNoisy, 2,880 callsNoisy, 1 s
Online8431111
Raw Ritz8341010
NoFallback78043
Fixed one-edge oracle30011
NAGLA H2 (tuned)76777
RestartNAG85777
  • Rollback (Online vs. NoFallback). On historical starts at 360 calls, rollback rescues 6 starts with no losses (84 vs. 78; +4.17 pp, 95% file-cluster CI [1.39,7.64][1.39, 7.64]). On noisy starts it gives 11 vs. 4 at 2,880 calls (seven rescues, no losses; +29.17 pp, CI [12.5,50][12.5, 50]) and 11 vs. 3 at 1 s.
  • Spectral estimate (Online vs. Raw Ritz). Replacing the running maximum by the latest estimate changes little (83 vs. 84; 10 vs. 11), so we do not claim a large isolated benefit from the envelope.
  • Fixed oracle. Despite its optimal local factor, the fixed one-edge oracle reaches only 30/144 and 1/24. Local optimality need not translate into finite-budget recovery from nonlocal starts.
  • Tuned baselines. On noisy starts Online trails at 360 calls (3 vs. 7) but reaches 11 vs. 7 at 2,880 calls and at 1 s (+16.67 pp, CI [0,37.5][0, 37.5]). These CIs are exploratory and n=24n=24 is small.

Connection to the theory (Table 2). Only 1.40% (99/7,072) of early synthesized edges (within 360 calls) satisfy the tolerance of Theorem 4.2, and 108/192 trajectories reach the ceiling estimate. In the near-orbit subset (eX≤10−6e_X\le 10^{-6}), compatibility rises to 81.31%, and the fallback fraction drops from 1.70% overall to 0.055% (8/14,458). These descriptive diagnostics are consistent with Theorems 4.2 and 5.1: far from the orbit, estimates rarely meet the required gap-scale accuracy and the safeguard of Theorem 6.1 controls progress; near the orbit, estimates are more often acceleration-compatible. They do not establish a causal effect, and no trajectory satisfies the conservative certified-tube proxy of Theorem 6.1. We therefore present robustness and local acceleration as complementary results, not as one implying the other.

Comments on the paper

1 comment

Dear Reviewers,

Now that the authors have responded, please review what they have said and give consideration to whether they have adequately addressed your concerns.

-The AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.