acceptodds
Withdrawn from ICLR 2027

Diagonal-Preserving Power-Coherence Parameterization for Learned Covariance Source Imaging

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learned estimators for coherent acoustic imaging predict a source power map from microphone cross-spectral data, but the measurements depend on the full source covariance, of which the map is only the diagonal. We study how a learned estimate should be refined against the measurements within a fixed optimization budget, and show that the choice of covariance coordinates matters. The conventional diagonal-plus-low-rank form couples power and coherence, so updating its coherent factor also changes the map. We adapt a bounded factor-correlation model to complex covariance: the map remains the covariance diagonal, and positive semidefiniteness holds without a dense projection. Both forms describe the same trace-normalized covariance family and can start from the same physical covariance, so matched controls isolate the effect of coordinates from learned initialization and learning rate selection. With separate development-selected rates for power and factors, power–coherence (PC) refinement lowers shape error by 26.5–30.0% and raises spatial overlap by 11.3–14.6 percentage points after 80 updates without a map anchor, across eight paired seeds and three initializations. The largest gains from CNN starts occur on coherent lines and arcs, with consistent gains on synthetic blade mixtures. With an anchor, PC lowers the joint objective by 17.4–18.6%. An independent time-domain simulator also shows spatial gains without retuning, while Gaussian and Fourier operators show objective reductions. These results identify optimization coordinates as a design choice for refining learned estimates in covariance inverse problems, even when the conventional form attains a lower measurement residual.

Primary area:
optimization
Keywords:
covariance parameterization, finite-budget optimization, inverse problems, learned initialization, acoustic source imaging

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_jFQQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how to refine a learned estimate of a coherent acoustic source covariance against the microphone cross-spectral matrix (CSM) within a fixed optimization budget. In the usual diagonal-plus-low-rank form, updating the coherent factor also changes the source power map. The authors propose a power–coherence (PC) parameterization. This is a complex-valued version of Borsdorf et al. (2010). In PC, the map is exactly the covariance diagonal, PSD holds by construction, and projection costs the same as the conventional form. Both forms span the same trace-normalized covariance family, so any differences come from the optimization path. A small CNN initializes power and coherence from beamforming maps, followed by 80 Adam refinement steps with an optional map anchor. Validation is mainly on synthetic analytic CSMs (21×17 grid, 32-microphone circular and spiral arrays, six source families), comparing PC against the conventional form from matched initial covariances with separately tuned rates. Without an anchor, PC reduces shape error by 26.5–30.0% and raises IoU by 11.3–14.6 points, with the largest gains on coherent lines, arcs, and blade mixtures. Conventional keeps a lower CSM residual. With an anchor, PC lowers the joint objective by 17.4–18.6%. Further experiments cover an Acoular time-domain simulator (small IoU gains, both methods with low absolute overlap), Gaussian and Fourier operators (objective reductions only), comparisons with classical and learned baselines (BF, DAMAS, CLEAN-SC, covariance-fitting variants, DAMAS-FISTA-Net), and regularized fitting of wind-turbine proxy CSMs without ground truth.

Critical strengths

None

Critical weaknesses
  1. Limited methodological novelty. The core construction, S = D_p^{1/2} R(V) D_p^{1/2} with a unit-diagonal factor-correlation matrix, is by the authors' own account (Sec. 4.1) a complex-valued version of the real factor-correlation model of Borsdorf et al. (2010), and the scale/correlation separation goes back to Barnard et al. (2000). Extending it to Hermitian matrices is straightforward, and the PSD/diagonal properties (Eqs. 3–4) and Proposition 1 follow directly from the construction. In addition, the related-work section omits learned covariance/CSM estimation work relevant to the positioning, e.g., SubspaceNet (learned surrogate covariances for coherent sources in DoA estimation), neural-network covariance reconstruction for DoA, GAN-based CSM filtering, and CNN-based microphone covariance upsampling for acoustic imaging. This matters because the claimed novelty is the coordinate choice, and the paper neither shows that this is non-obvious nor positions it against learned covariance-estimation approaches.

Shmuel, Dor Haim, et al. "SubspaceNet: Deep learning-aided subspace methods for DoA estimation." IEEE Transactions on Vehicular Technology 74.3 (2024): 4962-4976.

A. Barthelme, W. Utschick, "DoA Estimation Using Neural Network-Based Covariance Matrix Reconstruction," IEEE Signal Processing Letters,

Puhle, Christof. "Deep learning-based filtering of cross-spectral matrices using generative adversarial networks." arXiv preprint arXiv:2502.21097 (2025).

  1. Insufficient experimental comparison. The main evidence (Tables 1–3) compares PC against the authors' own Conventional parameterization, and the paper itself states that the external comparisons only complement these matched controls (Sec. 5.5, App. D.1). Established coherent/correlated-source methods are not compared, although some are cited: MACS (the covariance-fitting baselines are explicitly not reproductions of MACS), generalized inverse beamforming, DAMAS-C, SC-DAMAS, CMF-C, or SBL-type methods. Most baselines (BF, DAMAS-PGD, Sparse-/Reweighted-CF) assume incoherent sources, while the claimed gains are concentrated on coherent lines. No learned method that models coherence is compared.

  2. The paper requires significant writing improvements. Its narrative is not consistent; the abstract mentions “We study how a learned estimate should be refined against the measurements within a fixed optimization budget, and show that the choice of covariance coordinates matters”, but the method section does not involve any budget parameter or setting. Overall, it is not clear what the main motivation of the paper is compared with previous approaches.

Questions for authors
  1. Beyond extending Borsdorf et al. (2010) to complex Hermitian matrices, what is technically new in the PC parameterization? Is there any non-trivial difficulty in the complex case?

  2. Since both forms represent the same covariance family, do the spatial gains persist when both refiners are run to convergence, or with stronger optimizers (e.g., L-BFGS, or more than 320 updates with tuned rates)? If the gap closes, the contribution is mainly about a fixed 80-step budget.

  3. Could a simpler fix to the Conventional form, such as projecting factor updates to preserve the diagonal or reparameterizing p as the full diagonal, recover most of the PC gain?

  4. How was the factor rank r = 6 chosen? Appendix C states only that PC lowers the anchored objective at ranks 1, 2, 4, 6 and 12, and notes a reversal at rank 12 after 20 updates. Please report shape-L1, IoU, FP-pix, recall and CSM error as a function of rank in the unanchored two-rate setting (Table 2), ideally by source family. Does the PC advantage shrink as r grows relative to the true rank of the coherent components? The main claim is about spatial quality, so a rank sensitivity reported only for the objective is not sufficient.

  5. How does PC compare with established coherent-source methods (MACS as originally proposed, generalized inverse beamforming, DAMAS-C, SC-DAMAS/CMF) and with learned methods that model coherence or estimate covariances (e.g., SubspaceNet-style surrogate covariance estimators)?

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

The paper proposes a power-coherence (PC) approach for solving the inverse problem in acoustic source imaging. In the conventional method, updating the coherence factors would change the power map due to the coupled modeling. In the proposed PC method, the diagonals are preserved by using separate power and coherence coordinates. This allows the update of the coherence factors independent of the power map. As a result, it solves the coupling flaw in the conventional method. Experimental results show that the proposed PC approach improves spatial shape estimates compared to the conventional method.

Critical strengths

The experimental results (Table 2) show that the PC method reduces the shape error by 26.6% or up to 30%. These improvements seem to be significant although the acoustic benchmark is synthetic. However, it is not clear whether the baseline is the current SOTA.

Critical weaknesses

I find it really hard to follow the paper. For example, in Section 3, the problem formulation is not clear enough. A lot of background materials and definitions are missing. For example, the definition of "CSM" is missing in the paper. The mathematical definition of the steering matrix HH is missing. What is the distribution of the noise NN? It is not clear how the problem is set up and why the conventional method has to be the baseline. It is unclear whether the coupled approach is the SOTA approach that one should compare to. On the other hand, in terms of mathematical rigor, Proposition 1 seems redundant and trivial given the setup and it is not clear whether the PC approach is optimal. It seems ad-hoc and there could be many similar formulations leading to the same decoupled estimation of power and coherence. A discussion can be insightful to understand why this particular form is chosen.

Questions for authors
  1. What is the definition of a beamforming map?
  2. Is the PC method optimal in certain sense? There is usually some optimality in methods solving inverse problems.
  3. How to determine the best anchor strength λ\lambda?
  4. What is the design motivation of LCSML_{CSM}? It seems scale-invariant due to the free choice of β\beta.
  5. Is the coupled method (the conventional method) the strongest baseline known?
  6. Could you explain more on why updating LL would impact the estimation performance on the power? If a model is optimized to jointly estimate pp and LL, does it make sense that LL should learn to compensate based on pp? Is there a proof for the conventional method being suboptimal?
Reviewer_nVgLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the parameterization of covariance matrices in learning-based acoustic source imaging. To address the issue of coupling between source power and spatial coherence—a common drawback in traditional formulations that represent the covariance matrix as a diagonal matrix plus a low-rank factor—the authors propose a "Power–Coherence" (PC) parameterization. By employing a bounded factor-correlation structure, this approach ensures that the covariance matrix remains positive semi-definite while keeping its diagonal elements strictly aligned with the source power map, thereby achieving an explicit separation of power and coherence. The authors further demonstrate that, following trace normalization, the PC parameterization and traditional parameterizations represent the same set of covariance matrices; the primary contribution lies in the transformation of the optimization coordinates and the associated finite-step optimization process. The proposed method is systematically evaluated across six types of synthetic source distributions, two microphone array geometries, and various initialization conditions. The results indicate that the method offers the most significant advantages for scenarios involving highly coherent sources, such as those arranged in lines, arcs, or complex, multi-lobed configurations. Furthermore, the approach is validated through experiments using Acoular time-domain simulations, Gaussian and Fourier observation operators, and regularized fitting based on real-world measurement data from wind turbines.

Critical strengths
  1. The PC parameterization proposed in this paper explicitly decouples source power from spatial coherence. Through a structured formulation, it ensures that the covariance matrix remains positive semi-definite and that the diagonal elements—representing source powers—are consistent. Rather than seeking performance gains by expanding the model’s expressive capacity, the authors maintain the same covariance-modeling capability as traditional parameterizations while investigating how different parameterization structures influence the finite-step optimization process. This offers a valuable perspective for understanding how the structure of a parameterization affects covariance optimization and spatial reconstruction.
  2. The experimental design is systematic, and the performance trade-offs of the proposed method are analyzed in considerable detail. The authors conducted paired comparative experiments across a variety of source configurations, array geometries, and initialization conditions.
Critical weaknesses

1: Lack of validation using sufficient real-world data; reliance on simulated data distributions may limit the external validity of the conclusions. 2: The optimization advantages of PC (Parameter-Conditioned/Parametric-Coordinate) methods are not sufficiently explained, making it difficult to distinguish between the benefits of parameterized geometry and the coordinate-dependent effects inherent in the Adam optimizer. 3: The advantages of the proposed method are highly dependent on the coherence of the sound sources; its scope of applicability and the breadth of experimental coverage have not been adequately demonstrated.

Questions for authors

First, all six types of sound sources in your study are derived from your own simulated experiments; does this imply that the results might be influenced by the specific, artificially constructed data distribution? Furthermore, why did you not utilize publicly available datasets—such as MIRACLE—that would be well-suited for your research? While your validation using synthetic data is quite thorough, validation based on real-world propagation is relatively limited, and validation in practical applications lacks reliable ground-truth localization data. Second, you have demonstrated that both the PC (parameter-coordinate) approach and the conventional approach can represent the same set of covariance matrices; however, the advantage of the PC approach becomes apparent primarily when the number of optimization steps is limited. If the conventional approach were granted a sufficient number of optimization steps, could it eventually achieve the same spatial reconstruction results as the PC approach? Additionally, is the advantage of the PC approach derived from a more sound optimization geometry, or is it simply a consequence of the Adam optimizer producing different update scales when applied to the two different parameter coordinate systems? It appears there are no experiments to substantiate this. Third, Table 6 indicates that the advantage of the PC approach is most pronounced for distributed sound sources with high coherence. Why is this advantage concentrated specifically on coherent linear and arc-shaped sources, rather than on standard point sources or general linear sources? Does this suggest that the coupling between power and coherence—inherent in conventional parameterizations—exerts a more severe negative impact on spatial reconstruction only when significant spatial coherence is present among the sound sources? This also points to potential issues regarding the completeness of your experimental datasets.

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

The paper proposes a power-coherence parameterization for acoustic source covariance that preserves the power-map diagonal during coherence updates, maintains positive semidefiniteness, and represents the same trace-normalized covariance family as the conventional formulation. Matched experiments demonstrate improved synthetic spatial reconstruction under fixed optimization budgets, although the benefits vary with the source family and refinement objective (lines 303-322 and 358-362).

Critical strengths
  1. The experimental controls are thorough. Both parameterizations share the initial covariance, rank, objective, and update budget (lines 221-223), and receive symmetric searches over separate power and factor learning rates (lines 295-299). Consistent spatial gains across eight paired seeds and both CNN initialization origins strengthen the evidence for a coordinate effect (lines 309-313).

  2. The formulation is simple and computationally practical, preserving positive semidefiniteness without forming a dense source covariance and retaining the conventional formulation's asymptotic projection complexity (lines 172-178). The synthetic gains are substantial, with shape-error reductions of 26.5% to 30.0% under the separately tuned, unanchored comparison (lines 303-304).

Critical weaknesses
  1. The methodological novelty is limited. The construction adapts the established factor-correlation parameterization of Borsdorf et al. (2010) to complex covariances, as acknowledged in lines 164-165. The Jacobian analysis explains differences in optimization trajectories but does not identify conditions under which they improve source recovery (lines 212-220), leaving the broader significance primarily dependent on empirical evidence.

  2. Evidence of practical reconstruction quality remains limited. Although PC improves over the conventional baseline in the independent simulator, absolute IoU remains between 0.044 and 0.047 (lines 416-419), warranting additional localization and threshold-sensitivity analysis. The measured-data experiments use proxy covariances without localization ground truth (lines 429-431 and 462-463). Furthermore, the weak-source injection experiment reports 9.3% success for PC versus 63.0% for the conventional method under the matched, -12 dB, eigenvalue-clipping condition with an anchor (Table 16; lines 1354-1365).

Questions for authors
  1. What causes the low absolute overlap in the Acoular experiments reported in lines 416-419? Do peak-localization error, recovered source power, and support-threshold sensitivity indicate useful reconstruction despite the low IoU?

  2. Given the proxy-covariance construction and absence of localization ground truth acknowledged in lines 429-431 and 462-463, can the authors provide validation using synchronous microphone recordings with known source positions?

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

Related papers

Discussion (0)

Sign in to comment.