acceptodds
Under review as a conference paper at ICLR 2027

QuIR: Quotient-Informed Recovery Network for Joint Sparse Blind Deconvolution

Abstract

Jointly sparse blind deconvolution (JSBD) seeks to recover multiple signals and a shared unknown convolution kernel from noisy observations, assuming that the signals have a common sparse support. Existing inverse-filter methods apply a candidate inverse filter to the observations and evaluate the joint sparsity of the resulting signals. This operation can amplify measurement noise at frequencies where the kernel response is weak, degrading recovery under severe ill-conditioning or very low signal-to-noise ratios. We propose Quotient-Informed Recovery Network (QuIR), which predicts a candidate support before inverse-filter estimation to guide signal and kernel recovery. To predict support from a limited number of noisy channels, we train a neural network to estimate normalized counts of cyclic coordinate differences between support locations. The network uses quotient features that do not depend on the unknown kernel in the noiseless model. On the evaluated noise-conditioning grid, QuIR achieves exact support recovery, compared with for inverse-filter JSBD.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Structured representation learning, Jointly sparse blind deconvolution, Model-based recovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fZbzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies sparse signal support recovery in the problem of jointly sparse blind deconvolution and proposes QuIR, i.e., Quotient-Informed Recovery Network. The framework contains three stages. In the first stage, the paper identifies the shift ambiguity in recovery and hence focuses on the support’s circular autocorrelation, essentially recording the differences between nonzero positions. Further, under several assumptions, the paper shows that this normalized circular autocorrelation can be obtained in the large-channel limit by applying an inverse Fourier transform and zero-offset normalization to a sequence computed from averaged squared magnitudes of normalized inner products between Fourier rows of the original measurements. However, finite-channel effects and noise can make this direct estimate inaccurate. Instead of directly applying this relation, the paper proposes a learning-based method, followed by a non-learned decoder to translate the estimated circular autocorrelation into support candidates, which constitutes Stage II. In Stage III, the relative off-support-energy cost in (3) is written as a generalized Rayleigh quotient and hence solved through a generalized eigenvalue problem. The paper also identifies the sources of error and presents conditional theoretical guarantees concerning their effects on recovery. Finally, evaluations on synthetic and measured data demonstrate the efficacy of the proposed QuIR.

Critical strengths
  1. Coherent structure design in QuIR. For each individual stage, its role is clear and the choices are also well motivated both by intuition and derivation. Built upon three separate stages, QuIR is also well connected and the complete roadmap is promising.

  2. The analysis in sec. 3 explains conditions for resolving reflection from the original observations and for obtaining an accurate inverse direction from an inexact candidate. These results are insightful and help interpret the method, even though the guarantees are conditional and some proof ingredients are standard.

  3. The paper also provides informative experiments and reports not only the success and improvement but also its limitations.

Critical weaknesses

None

Questions for authors
  1. I would like to mention Sensor Calibration for Off-the-Grid Spectral Estimation by Eldar, Liao, and Tang. After Fourier transformation, QuIR’s model (1) can be viewed as a fixed-grid instance of their diagonal-calibration model, up to normalization. Their covariance-based methods retain complex phase information and exploit the underlying Toeplitz structure to estimate unknown calibration gains and phases, which correspond to QuIR’s kernel frequency response \hat g. Under their stated conditions, they can also recover the spectral support up to a common shift. QuIR takes a different route. Its squared magnitude eliminate dependence on the kernel phase. The resulting autocorrelation representation consequently retains reflection and possible ambiguities. Stage III then uses the original observations to resolve such ambiguities and estimate an inverse. I would like to know the potential advantages and limitations of QuIR over covariance-based self-calibration. In particular, it would be useful to clarify whether avoiding explicit phase calibration offers benefits in terms of, for example, few channels, noisy observations, or poor kernel conditions.

  2. Since the candidate search already considers multiple cardinalities, could the authors discuss how QuIR’s outputs could be used to estimate K for the final top K support extraction when the true cardinality is unavailable?

  3. A waveform NMSE of 0.375 is reported. Could the authors discuss the main sources of this remaining waveform error?

5 comments
AuthorsOpenReview ↗

Clarifying QuIR’s Support Estimation, Cardinality Readout, and Reconstruction Accuracy

We thank the reviewer for prompting us to clarify three points: QuIR's relationship to covariance-based self-calibration, how the decoded candidate supports readout when the support size is unknown, and why exact support recovery can coexist with waveform error. We address these points in turn.

AuthorsOpenReview ↗

1. Relation to covariance-based self-calibration

QuIR and covariance-based self-calibration share a diagonal-gain formulation on a fixed Fourier grid, but exploit different structural information. In the noiseless setting, QuIR's observation model is Y^=diag⁡(g^)FSXS\widehat{Y}=\operatorname{diag}(\widehat{g})F_SX_S, where Y^\widehat{Y} is the Fourier-domain observation matrix, g^\widehat{g} is the kernel frequency response, FSF_S contains the Fourier columns indexed by the common support SS, and XSX_S contains the active coefficients across channels. The kernel response therefore plays a role analogous to the diagonal calibration gains in the model of Eldar, Liao, and Tang. Their method retains the complex covariance and exploits the Toeplitz structure of the underlying covariance to estimate the calibration gains and spectral support. QuIR instead forms normalized comparisons of Fourier rows. Under the noiseless independent-coefficient model, these comparisons remove each row's nonzero complex scaling and yield the normalized support autocorrelation in the large-channel limit. QuIR then uses the original complex observations to compare the decoded candidate with its reflection and fit the filter.

Because the support-autocorrelation estimate encodes relative spacings, it is invariant to reflection and may also be shared by homometric supports. QuIR uses the original observations to compare the decoded candidate with its reflection, but other homometric alternatives may not be tested. With few channels, the empirical row comparisons may deviate from their large-channel limit. Under noise, row directions are less reliable at frequencies where the kernel response is weak. QuIR therefore also uses row-power features, which provide information about observed signal strength but remain dependent on the kernel response. Our experiments evaluate QuIR under these conditions; they do not isolate the effect of avoiding explicit phase calibration or establish a performance advantage over covariance-based calibration at matched channel counts, noise levels, and kernel conditioning. Section 1 discusses this relationship.

AuthorsOpenReview ↗

2. Candidate-based support-size readout

We evaluated the candidate-size readout on 114,200 recorded synthetic scenes across seven sweeps, holding the full reconstruction and row-energy ranking fixed. The true-KK readout selects the KK highest-energy rows and serves as a reference for evaluating support localization when the size is known; the reconstruction itself does not use the true KK. The candidate-size readout selects the K^cand\widehat{K}_{\mathrm{cand}} highest-energy rows. With the ranking fixed, it recovers the exact support up to circular shift if and only if the candidate size equals KK and the true-KK readout succeeds. We used this rule to compute the exact-recovery counts from the recorded true sizes, candidate sizes, and true-KK outcomes. The archive QuIR_estimated_K_readout.zip, included in QuIR_reviewer_supplement.zip, contains the per-scene records (frozen_readout_records.csv), the summary (summary.json), and the script used to recompute the counts (reproduce_counts.py).

The candidate size matches KK in 86,608 scenes (75.84%). Exact support recovery up to circular shift is 75.58% (86,307/114,200) with the candidate-size readout and 87.32% (99,715/114,200) with the true-KK readout. Among scenes with the correct candidate size, 86,307 of 86,608 (99.65%) yield exact recovery. In 13,408 scenes with an incorrect candidate size, selecting the true number of highest-energy rows from the same reconstruction still recovers the exact support. In those scenes, the estimated size causes the readout to fail even though the row-energy ranking is sufficient to localize the support given KK. The effect is particularly pronounced in the low-channel signal-length sweep, where exact recovery decreases from 87.05% to 43.83%.

Reviewer_Hbt4OpenReview ↗
Rating
1
Confidence
1(−1)
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper claims to have found a novel method for joint-sparse blind deconvolution. The latter problem refers to recovering a filter and a number of sparse vectors xℓx_\ell from observations yℓ=g∗xℓ+wℓy_\ell=g*x_\ell+w_\ell, where the xℓx_\ell are assumed to be jointly sparse, i.e., suppxℓ\mathrm{supp} x_\ell is the same for all ℓ\ell.

The method is admittedly somewhat hard to decipher, but the main idea (as presented in figure 1) seems to be the following: The support of the xℓx_\ell can be recovered using a normalized support autocorrelation aa ( the autocorrelation of the vector 1\suppxℓ1_{\supp x_\ell}. The authors build a neural network to estimate aa from hand-crafted features of the yy-vectors, and then turn their estimate into an actual support-autocorrelation vector. Once the support has been recovered, it is not hard to reconstruct the xℓx_\ell.

The validation is likewise hard to decipher. They seem to evaluate their methods by solving angle-of-arrival problems, and microscopy imaging problems form microscopy.

Critical strengths

The empirical results seem good.

Critical weaknesses

The manuscript is hard to read. Despite careful reading and inspection of the code, I can only approximately pinpoint the authors' approach. There is no single algorithm presented, and most statements are not formally presented as theorems, so that their proofs cannot easily be found in the appendix. When propositional statements are defined, their proofs are not clearly referenced in the appendix. There are also many seemingly non-standard terms that are not defined in the manuscript. See also questions below.

Questions for authors

My assessment very much comes down to me not being able to parse many details in the paper. Here are a list of detailed concerns.

  1. What are the exact connections to the referenced manuscript (Wang et. al,2026)? The authors seem to claim that said paper bears the most resemblance to their manuscript, but the connections are not made clear.

  2. What are the vectors ξk,ℓ\xi_{k,\ell} in section A.2. in the appendix?

  3. Can the authors explain more thoroughly how they solve the problem arg⁡min⁡T∈Scand∣∣aT−a^∣∣22\arg \min_{T\in \mathcal{S}_\mathrm{cand}}||a_T-\widehat{a}||_2^2, and in particular how they determine Scand\mathcal{S}_{\mathrm{cand}}? In particular, in D.3., what is the explicitly seeded start and why is it used? What are local swaps? They mention using an eigengap as a deterministic tie-breaker? Why, and eigengap of what?

  4. What is the meaning of 'homometric supports'?

  5. In the end of section 2, the authors say that the 'true scene cardinality is supplied' for the final readout. How is the true scene cardinality defiend, and why is it sound to supply it for the final readout?

  6. Can the authors motivate their feature construction in appendix D.1. better? Why is R(v)R(v) defined via dividing another quantity by 1.4826 and thereafter clipping to [−20,20][-20,20]? Why are row powers used? What does it mean to repeat descriptions 'across lag'?

  7. What are 'hybrid circular/spectral blocks' in appendix D.2?

  8. In the same section., the authors say that 'checkpoint selection uses held-out MSE between the projected projection a^\widehat{a} and aSa_S': What does this mean?

  9. Again in the same section, the authors claim that they are using a 240/60 train/validation split. Is this a typo? If not, what does this mean?

  10. Again in the same section ,the authors claim that they 'transfer 38 compatible tensors from a separately trained model'. Why 38? What is the (architecture) of the other model?

  11. In the evaluation section, the setup can be described more thoroughly. What are scenes and cells? What is the meaning of 'indoor location' at the end of page 8? Is the same data used for all experiments, or does it vary between experiments?

  12. Again in the evaluation section, what is the meaning of a 'preserved archive on matched grids, so that pooled comparison is unpaired'?

  13. How is recovery of a scene defined in the evaluation section?

  14. What is the importance of the results presented in A.4 and A.5, and why are they not referenced in the main paper?

  15. What are the 'ideal' and 'deployed banks' in Section B.3? In the same section, what is the map N\mathcal{N} in the definition of the 'measurable comparator'? What does the 'orthogonal decomposition above' refer to?

  16. In B.4., what does it mean that 'a zero margin prevents the argument from distinguishing a homometric competitor'?

  17. In proposition 2, how is a 'strictly positive clean inverse filter residual' defined? What are conditions for the clean power normalization to be positive definite? In its proof in Section C.1., what are the matrices DℓD_\ell in Section C.1?

  18. What is the statement shown in C.3?

  19. What is a 'phase-ramp gauge' in C.4?

  20. In table 2, what is an 'intermediate-KK stress inside the declared training scope'? What is the meaning of 'phase axis'? And what do the numbers mean?

  21. What do the authors mean in their AI use statement when they state that the 'numerical claims were computed from preserved fixed-protocol result files?'

10 comments
AuthorsOpenReview ↗

Clarifying QuIR’s Methodology, Theoretical Guarantees, and Empirical Evaluation

We thank the reviewer for carefully reading the manuscript and code. Your summary correctly identifies the first two stages of QuIR: estimating the normalized support autocorrelation and decoding it into a candidate support. We agree that the original manuscript did not explain clearly how this candidate is used in the final reconstruction. The revised manuscript now presents the full procedure in Figure 1 and Algorithm 1: the candidate specifies where the recovered signals are encouraged to concentrate their energy; QuIR fits the corresponding filter to the original observations and retains the reconstructed signal at every location. Thus, the final support can differ from the decoded candidate. We have also clarified the notation, experimental protocols, and the theoretical proofs. We organize our responses below around the method, its analysis, and its evaluation.

AuthorsOpenReview ↗

1. QuIR's methodology and its relation to prior work.

QuIR addresses jointly sparse blind deconvolution (JSBD) under a multichannel model in which each latent signal is circularly convolved with a shared unknown kernel and corrupted by additive noise; in matrix form, the observations satisfy Y=C(g)X+WY=C(g)X+W. The rows of XX share a common sparse support. QuIR aims to recover this support structure and the latent signals without requiring the scene-specific support size. Its framework has three stages: it first constructs support-informed features from the observations, then estimates the normalized support autocorrelation and decodes a candidate support, and finally uses that candidate to guide filter fitting on the original observations. The reconstruction retains every location, so the final support can differ from the decoded candidate.

The first stage uses Fourier-row comparisons to capture support-related structure while accounting for noise. Let Y^\widehat{Y} be the columnwise discrete Fourier transform of YY. The vector rk=(Y^k,:)Tr_k=(\widehat{Y}_{k,:})^{\mathsf T} collects the observed coefficients across channels at frequency kk, and its squared norm, pk=∥rk∥22p_k=\|r_k\|_2^2, is the total row power. The latent Fourier row is defined analogously. In the noiseless model, the kernel multiplies each Fourier row by a frequency-dependent complex scalar, which normalized row comparisons remove when the rows and kernel responses are nonzero. Under additive noise, row directions become less reliable when the signal component is weak. QuIR therefore combines unweighted comparisons with comparisons that downweight low-power rows and with summaries of the observed row powers. These power-based features depend on the kernel. Section 2.1 motivates this design, and Appendix A.2 provides the formulas.

In the second stage, a network maps the features to a score vector and projects it onto the set of admissible normalized support autocorrelations. During supervised training, the known support supplies the target autocorrelation. We use the quadratic Fenchel–Young loss because its gradient with respect to the scores is proportional to the difference between the projected prediction and the target. With the normalization in Section 2.2, the loss also bounds the squared representation error. The decoder then compares each candidate's normalized autocorrelation with the network estimate, using their squared Euclidean distance. For each allowed support size, it fixes one location at zero because circular shifts preserve autocorrelation. It uses six deterministic initializations based on the largest predicted values at nonzero lags and one weighted random initialization controlled by a fixed seed. Each initialization is greedily expanded to the target size. A local swap replaces one selected nonzero location with an unselected location. The decoder applies the best swap whose mismatch reduction exceeds the numerical tolerance and stops after at most three passes or when no improving swap remains. It returns the candidate with the smallest mismatch across the prescribed size range. Appendix A.4 specifies the search and stopping rules.

In the third stage, the decoded candidate defines where the recovered signals are encouraged to concentrate their energy. For both the candidate and its reflection, the solver fits a filter to the original observations by minimizing the fraction of reconstructed energy outside the corresponding support. It selects the orientation with the smaller residual. If the residuals tie, it uses the eigengap: the difference between the two smallest eigenvalues of the normalized filter objective matrix. This gap measures the separation between the minimizing eigendirection and the next eigendirection and serves only as a deterministic tie-breaker. Sections 2.3 and Appendix A.5 define the matrix and its equivalent Gram computation. QuIR then applies the selected filter to the observations and retains the full signal estimate.

It is noteworthy that the inference procedure does not use the scene-specific true support size, although the decoder searches a prescribed range of candidate sizes and reports the size of its candidate. In the original synthetic support metric, the true support size is used only after reconstruction to select that many rows with the largest energies. This readout evaluates localization when the support size is known; it does not measure final support recovery when the size must also be estimated. Section 4 reports the results obtained when the candidate size is used for the final readout. The implementation details follow the three-stage design. The feature tensor contains 16 lag maps and 12 global scalars, with each global scalar repeated across all lags. For robust rescaling, each lag map is centered by its median over nonzero lags, divided by 1.4826 times its median absolute deviation plus 10−910^{-9}, and clipped to [−20,20][-20,20]. The factor 1.4826 scales the median absolute deviation to the standard-deviation scale under a Gaussian reference. The resulting 28×N28\times N tensor is processed by a network with width 36, three blocks, and 24 spectral modes per block. Each block combines circular convolution across neighboring lags, spectral convolution on the Fourier modes of the lag sequence, and pointwise convolution across feature channels. A learned gate combines the circular and spectral outputs before the pointwise-convolution output is added. Appendix A.3 describes the architecture.

To train the support-autocorrelation estimator in QuIR, we use a two-stage procedure: an earlier model provides the initialization, and a final training stage adapts it to QuIR's inference features. The earlier model was pretrained on 240 synthetic scenes and validated on 60; the final model was trained on 800 scenes and validated on 160. Both models use the same width, number of spectral modes, and number of blocks. During transfer, 38 tensors were copied or remapped. Three global features in the earlier model depended directly on the true support size, so their input columns were omitted when adapting the model to the final feature set, while compatible downstream parameters were retained. Thus, the final inference interface does not require the scene-specific support size, although some transferred parameters were learned from a model whose inputs included support-size-dependent features. The final network is trained with the signal-length-normalized Fenchel–Young loss. We selected the checkpoint using the validation mean squared error between the projected prediction and the target autocorrelation, normalized by the signal length; the validation Fenchel–Young loss selected the same checkpoint. Appendix A.3 describes the training and checkpoint-selection procedures. To verify reproducibility, the supplementary package provides the historical initialization, data-reconstruction and training scripts, and checkpoint-comparison records. Replaying the eight pretraining epochs and four final-stage epochs produced maximum absolute parameter differences of 2.35×10−52.35\times10^{-5} and 3.87×10−73.87\times10^{-7}, respectively.

Among prior work on JSBD, Wang et al. (2026) is particularly relevant to QuIR. Both methods consider multichannel circular-convolution measurements generated by a shared unknown kernel, with a common sparse support. They differ, however, in how they use this support information. In the noiseless setting considered by Wang et al., an inverse filter is applied to promote joint row sparsity in the filtered signals. Under additive noise, inverse filtering before support estimation can alter the noise statistics and amplify noise at frequencies where the kernel response is weak. QuIR addresses this noisy setting by first estimating support structure from the observations and decoding a candidate, then using that candidate to define the filter-fitting objective on the original observations. Thus, QuIR also fits a filter in its final recovery stage, but selects it by minimizing the normalized energy outside the candidate rather than by imposing joint row sparsity on the filtered outputs. Section 1, Figure 1, and Algorithm 1 clarify this relationship and the complete pipeline.

Reviewer_cFtAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

QuIR tries to address jointly sparse blind deconvolution by learning support autocorrelation from the statistics that are kernel-invariant in the noiseless setting. The approach then decodes this representation into a candidate support, which guides inverse-filter recovery. The study provides theoretical analysis and evaluates the approach on synthetic, wireless, and microscopy data.

Critical strengths

QuIR tries to recover signals from only a few observation channels, which is a challenging setting for existing blind-deconvolution methods. In the two reported low-channel evaluations, it improves exact support recovery over the cardinality-informed baseline by 12.33 and 15.24 percentage points. These gains demonstrate practical value of the approach and support future studies of using learned structural information for recovery when observations are limited.

Critical weaknesses
  1. The source of the performance gains is not clear. In table 1, the cardinality-informed baseline receives the true (K) during structural estimation, and the author state that "other pipeline components also differ" without clearly specifying those differences, which limits the attribution of the reported gains. Appendix E.4 provides a more direct comparison between support recovery using the learned autocorrelation estimate versus the analytic estimate while keeping the decoder and inverse-filter solver unchanged. However, the comparison was done on just 290 scenes. Extending this controlled comparison to the main evaluations would clarify how much of QuIR's improvement comes from learning the representation. Some design choices also provide limited benefit. Replacing projected MSE with the Fenchel-Young loss improves the final recovery only from 71.52 to 71.76%

  2. The theory can only partially explain the experiment results. Proposition 1, 3, and 4 provide conditional guarantees for support decoding, inverse-filter accuracy, and final support recovery. However, the experiments do not establish how often the sufficient conditions hold or whether the quantities in these bounds help explain the recovery failures. Thus, it remains unclear how much the analysis explains QuIR's performance with substantial noise or few observation channels.

Questions for authors
  1. Can you compare the learned and analytic representations across more challenging settings while keeping the rest of the recovery pipeline unchanged?
  2. Does the Fenchel-Young loss consistently improve the final recovery over projected MSE across repeated training runs?
  3. Can you clarify whether the theoretical conditions hold in practice and help explain successful and failed recovery?
  4. Can you clarify how does the final support recovery change when the number of active locations must be estimated instead of supplied?
  5. Given ground-truth autocorrelation, can the full pipeline distinguish supports with identical autocorrelations that are unrelated by shift or reflection?
7 comments
AuthorsOpenReview ↗

QuIR’s Recovery Mechanism: Methodological Connections, Empirical Evidence, and Limitations

We thank the reviewer for raising two central questions: how much of QuIR's recovery gain can be attributed to the learned autocorrelation estimate, and how far the theory explains the observed results. We distinguish complete-pipeline performance from controlled estimator attribution using 5,250 paired scenes across the main SNR–condition-number grid, assess the theoretical spectral quantities in a separate 290-scene diagnostic, and clarify the role of the training loss. We also clarify how QuIR handles the unknown support size and discuss the ambiguity caused by homometric supports. We organize our response around these three issues.

AuthorsOpenReview ↗

1. Attribution of recovery gains.

QuIR is designed to improve support recovery from noisy, finite-channel observations by learning a normalized support-autocorrelation estimate and using its decoded candidate to guide the subsequent filter and signal recovery. The theory describes a route by which this design can improve recovery under stated conditions: Proposition 1 relates autocorrelation-estimation error to candidate decoding; Proposition 3 bounds inverse-direction error in terms of candidate leakage and the eigengap; and Proposition 4 gives sufficient conditions for exact final support recovery. Thus, the proposed mechanism links improved support estimation to more accurate recovery, while the experiments below assess the size of the gain in practice.

The comparison with the cardinality-informed baseline measures the performance difference between the complete procedures. On the same 114,200 synthetic scenes, QuIR achieves 87.32% exact support recovery, compared with 84.04% for the baseline, a gain of 3.28 percentage points. The baseline uses the true support size KK both to estimate the representation and to perform fixed-size decoding. In its frozen implementation (in QuIR_cardinality_informed_baseline.zip; qisr_v3/v3_api.py, methods _branch and fit), it compares candidate orientations using a residual obtained by cross-fitting across channels, and selects the newer estimation branch only when both its autocorrelation mismatch and cross-fitted residual are lower than those of the earlier branch. QuIR instead searches over a prescribed range of candidate sizes, compares the decoded candidate with its reflection using the minimum normalized off-support energy on the original observations, and extracts the final support from the full reconstruction. The observed 3.28-point difference therefore reflects the complete pipelines and cannot, by itself, be attributed solely to the learned estimator. To isolate the estimator's contribution, we extended the fixed-downstream comparison to 5,250 paired scenes, with ten predetermined repetitions in each of the 525 SNR–condition-number settings of the main synthetic grid. We held the observations, seeded decoder, inverse solver, and final support readout fixed, changing only the supplied autocorrelation estimate. All three branches searched the same range of candidate sizes, while the true support size was used only for the final evaluation readout. Exact support recovery was 43.87% (2,303/5,250) with the original analytic estimate, 67.85% (3,562/5,250) after a simple median background correction, and 72.97% (3,831/5,250) with the learned estimate. The nonlearned correction therefore recovers much of the original 29.10-percentage-point gap, but the learned estimate retains a 5.12-point advantage over this stronger control. In the paired outcomes, 329 scenes were recovered only with the learned estimate and 60 only with the corrected analytic estimate. In the 560 scenes with SNR at or below −10-10 dB and kernel condition number at least 1,000, the original analytic, corrected analytic, and learned estimates recovered 24, 12, and 107 supports, respectively.

To examine whether this advantage merely reflects different candidate-size choices, we also fixed the decoder's candidate size to the true K=3K=3 in all three branches. Under this oracle-cardinality diagnostic, the original and corrected analytic estimates had identical exact-recovery outcomes at 52.90% (2,777/5,250), whereas the learned estimate achieved 63.37% (3,327/5,250). The 10.48-percentage-point gain shows that the learned estimate can improve recovery even when candidate size is fixed, although candidate locations and orientations may still differ between branches. These controls establish gains beyond this particular nonlearned correction and candidate-size selection, but do not exclude other nonlearned estimators or isolate individual network components. The scene-level paired outcomes, background-correction control, fixed-cardinality diagnostics, and reproduction scripts are provided in the supplementary mechanism-control archive (QUIR_MECHANISM_CONTROLS_5250_VERIFIED.zip within QuIR_reviewer_supplement.zip).

We use the quadratic Fenchel–Young loss to train the constrained autocorrelation estimator because its gradient with respect to the score vector is proportional to the difference between the projected prediction and the target autocorrelation. With the normalization in Section 2.2, the loss also upper-bounds the squared error of the projected prediction, aligning the training objective with the feasible representation used for support decoding. To test whether this property improves recovery, we compare the Fenchel–Young loss with projected mean squared error (MSE), holding the feature set and downstream procedure fixed. With one training seed, the two losses achieve exact recovery rates of 71.76% and 71.52%, respectively, on the same 2,100 scenes; on the noise subset, the corresponding rates are 54.75% and 55.17%. The overall difference is small, and the noise-subset result favors projected MSE. The ablation records are in original_evidence/09_QuIR_structural_ablation_6300.zip. These single-seed results do not establish a consistent recovery advantage for either loss across training runs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.