acceptodds
Under review as a conference paper at ICLR 2027

Missing-Data-Induced Phase Transitions in Spectral Partial Least Squares

Abstract

Spectral partial least squares (PLS-SVD) estimates the directions shared by two views of the same samples. We analyze it using a rank-one regression model, with entries of both views missing completely at random and filled with zeros. Missing response entries weaken the signal, whereas missing design entries also tilt the recovered direction. The setting is the proportional regime, with Gaussian response noise and a design that is column-orthogonal before masking. For incoherent designs and planted directions, we prove that as the signal diverges, the squared overlap with the truth approaches on the response side but only a ceiling on the design side. This ceiling is strictly below whenever design entries are missing, unless samples of vanishing leverage carry the signal. For this estimator, a stronger signal therefore cannot remove the design-side error. Under equal leverage and an explicit spectral condition on the masked design, which we prove for designs taken from nested Hadamard matrices, we establish a phase transition and derive the recovery curves. Under the same conditions, the critical signal strength is exactly proportional to , with the fraction of response entries retained. A heuristic replica derivation reproduces the equations behind these curves, and simulations follow them, including on random orthogonal designs outside the verified class. In semi-synthetic experiments on real biological designs with unequal leverage, the predicted ceiling varies across directions, and high-signal recovery tracks it. These results bring the theory of PLS-SVD closer to what is known for PCA under homogeneous missingness and show that the loss of signal-to-noise ratio observed there does not, by itself, describe missing design entries.

Primary area:
learning theory
Keywords:
random matrix theory, spiked models, phase transitions, partial least squares, missing data, high-dimensional statistics

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_ewDpOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper studies zero-filled PLS-SVD under MCAR missingness in a rank-one regression model with a column-orthogonal design. It derives the recovery threshold and singular-vector overlaps, and shows that missing design entries can cause a persistent recovery ceiling even when the signal becomes very strong. Simulations and semi-synthetic experiments on biological data are provided to examine these theoretical predictions. The manuscript and appendix together contain 73 pages, and I have not independently verified every theoretical statement and proof. Based on my overall reading, the main theoretical arguments appear reasonable.

Critical strengths
  1. The paper explains the different effects of missingness in the two views. Under the stated assumptions, missing response entries mainly weaken the signal, while missing design entries can cause a recovery ceiling even at high signal strength. This provides useful understanding of why stronger signals cannot always remove the design-side estimation error.
  2. The theoretical predictions are examined through simulations with different signal strengths, retention rates and sample sizes, together with semi-synthetic experiments on biological designs. These results provide useful support for the recovery curves and direction-specific ceilings within the studied model.
Critical weaknesses
  1. The theory is limited to a rank-one signal. Although eight biological directions are examined, each is planted and evaluated separately. This does not establish recovery when multiple shared factors coexist or interact. A rank-k analysis is still missing, which limits the breadth of the theoretical contribution.
  2. The practical applicability is not sufficiently demonstrated. The theory assumes homogeneous MCAR missingness and whitening before masking, while the biological experiments use synthetic responses and artificial MCAR masks. These experiments support the specified model, but provide limited evidence for naturally incomplete multimodal data.
Questions for authors
  1. Which parts of the analysis can extend to a rank-k signal, and which key steps specifically require the rank-one assumption?
  2. How can the direction-specific ceiling be used in practice when complete-data whitening and the true signal directions are unavailable?
  3. How does the Hadamard column-selection procedure in the code satisfy the nested-design condition of Corollary 3.6?
5 comments

Thank you for the careful review. Your Q3 caught an error in our code that we had missed. The figures and the text are corrected in the revision (changes in blue).

W1 and Q1. Rank one, and what extends to several components

You are right that the theory covers a single component and that the biological directions were planted one at a time. We do not have a theorem for several components. The new Appendix L.1 works out what carries over and tests two components planted together.

The high-signal results carry over. The decomposition in Proposition 3.1 is linear in the signal, so it holds exactly for each component. The geometry behind the ceiling (Proposition C.3) extends as well and brings in one new quantity, a cross coupling that measures how much the scores of two components share high-leverage samples. From these two results we compute the overlaps of all components at high signal. Under equal leverage the cross coupling vanishes, and components of distinct strengths each reach the same ceiling as when planted alone. A nonzero cross coupling mixes the components and changes their ceilings. In Table 4 the formula predicts design-side overlaps of 0.63 and 0.70, which is what we measure. Each component alone would reach 0.74. This is a computation and not a theorem, because we have not extended the perturbation argument behind Theorem 3.3.

Theorem 3.4, the finite-signal result, does need rank one, at two steps of its proof. The exact rank-one identity behind the threshold (Appendix E.6) gains terms between different components, and our spectral condition (Assumption E.10) gives no limits for them. The argument below the threshold (Appendix F.7) controls one component only. So for finite signal we report simulations under equal leverage (Table 5). Two components of distinct strengths each follow their own rank-one curve within about 0.01, and with equal strengths only their span is recovered.

On TCGA we planted the two leading directions together at high signal. With the response fully observed, each is recovered within 0.01 of its value when planted alone. With 30% of the response entries missing, the weaker one loses 0.08.

W2 and Q2. Using the results outside the model's assumptions

For the assumptions (new Appendix L.3, Table 7), we simulated masks whose retention differs across samples or features or depends on the value, and whitening applied after masking. The computation behind the ceiling, with a retention for each entry, predicts the high-signal design-side overlap within 0.02 in every setting. The assumptions do matter for the formula in the paper. At the mean retention it can be off by up to 0.45 when the design retention differs across features or depends on the value. The extension is a computation for the high-signal limit, not a theorem.

The true direction is no longer needed. The ceiling depends on it only through each sample's score along it, and these scores can be estimated from the response. With each sample's leverage estimated from its observed design entries, the new estimate of the ceiling uses only the masked data (Section 4, Appendix L.2). On the eight leading TCGA directions its mean over trials is within 0.005 of the true ceiling at high signal.

The estimate still assumes a design whitened before masking, which cannot be checked from the incomplete data. When we instead estimate the whitening from the masked data, recovery depends on where the small eigenvalues are cut off (Appendix L.3), so this remains open.

Naturally incomplete data also remain untested, because the true direction is then unknown. We tried the method on a real response from TCGA with an inconclusive result, described in our response to Reviewer TTS6 (Q1).

Q3. The Hadamard column selection in the code and the nested-design condition

The column selection in the code does not satisfy the condition. Corollary 3.6 covers the first DxD_x columns of the Sylvester matrix, up to signed permutations of rows and columns, while the released code draws DxD_x columns at random, which is in general a different design, and the submitted Appendix I.2 wrongly said that the corollary allows this. Figures 1(a) and 2 were therefore run on a design outside the verified class.

We re-ran both figures on the correct design, and the conclusions do not change. In Figure 1(a) the measured curves differ from the submitted ones by at most 0.012. In Figure 2 the values of Rx2R_x^2 differ by at most 0.043.

We have corrected the text in Appendix J.1 (I.2 in the submission) and at the end of Appendix G. The mistake concerns only the simulations behind these two figures, so no theorem or proof is affected.

We are happy to clarify anything further during the discussion, and we would be grateful if you would consider raising your score.

Thank you for the response and revision. The corrected Hadamard experiments and additional analyses strengthen the paper. However, (1) whitening from incomplete data remains unresolved, with recovery changing by 0.08–0.15 across the tested eigenvalue cutoffs; (2) the added experiments still use synthetic responses and controlled missingness, so validation on naturally incomplete data is still missing. These practical limitations and the incomplete rank-k theory restrict the scope of the contribution. I maintain my score of 3.

Thank you for the follow-up. We are glad that the corrected experiments and the new analyses helped and we agree with the limitations you list. On whitening, the loss comes from the smaller of the two cutoffs. At the larger one the overlap is at most 0.04 below what the complete-data whitener gives, but we have no rule for choosing the cutoff from the incomplete data, so this is open as you say. Incomplete data and a theorem for several components are also open and Appendix M lists all three. Some parts of the analysis hold for several components, and we will add a short paragraph to Section 4 so the main text reflects that.

Reviewer_TTS6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper provides an analytical characterization of the spectral partial least squares problem when solved using a rank-one regression model applied on masked observations under Gaussian noise. The summarized results are: (a) response masking attenuates the signal while design masking tilts the recovered direction, allowing for correct recovery of the response direction but limiting the correctness of the design direction, which is controlled by the design masking probability and the sample-to-feature ratio; (b) there is a critical signal strength under which the response and design direction reconstructions are misaligned.

Critical strengths

There are three characterizations of these results: theoretical results based on random matrix theory, numerical results that verify the theorems, and a replica analysis that provides agreement with the bounds without design assumptions. The characterization is asymptotic in nature and the theorem proofs are highly involved, and not easily checked due to their extent (60 pages). Nonetheless the numerical results are reassuring.

Critical weaknesses

The characterization is asymptotic in nature and the theorem proofs are highly involved, and not easily checked due to their extent (60 pages); the technical details in the appendices do often seep into the manuscript (e.g., descriptions of assumptions and results are brough up frequently).

There is little discussion of the practical applications of the rank-1 masked regression setup studied in the paper.

Questions for authors

Is it feasible to verify the regimes derived theoretically and checked numerically in an application setting with real world data? This would provide some reassurance that the rank-one setup that underlies the paper is useful in practice and that the results obtained are relevant in applied settings.

4 comments

Thank you for the good review and for the question about real data, which led us to test a real response. Changes in the revised PDF are in blue.

W1. Asymptotic results, long proofs, and technical detail in the main text

Asymptotic results. Yes, the two theorems are limits, as is standard for exact thresholds and overlaps of this kind. The decomposition in Proposition 3.1 is exact for every NN, and the ceiling in Theorem 3.3 is computed from the design at its actual size. The simulations show how close the limits are at finite size. In Figure 1(a), with N=2048N=2048, the predicted curves are within 0.005 of the measurements on average for signals at least 10% above the threshold. Figure 3(b) in Appendix J.2 shows the transition sharpening as NN grows from 100 to 5000. Appendices K.1 and K.2 report the noise settings where the predictions fail.

Technical detail in the main text. We moved several technical passages from Section 3 to the appendix, among them the detailed statement of the spectral condition, which Section 3 now only summarizes.

Checking the proofs. The proofs take 41 pages (Appendices B to G). The whole appendix was about 60 pages and also holds a reading guide, the additional experiments and the replica derivation. The two theorems also have separate proofs. Theorem 3.3, the ceiling, is proved in the first 8 of the 41 pages (Appendices B to D). Theorem 3.4 builds on the same setup and is proved in Appendices E and F, and Appendix G verifies its spectral condition for one class of designs. Appendix A is the reading guide, and its Table 1 lists what is proved for which class of designs. On why the proof of Theorem 3.4 cannot be made much shorter, please see our response to Reviewer B4ff (W2).

W2. Practical applications of the rank-one masked setup

You are right that we said little about practical applications, so the Introduction and the Conclusion now discuss them.

The setting we have in mind is two sets of measurements on the same samples, where we want the component they share and some entries are missing. Paired omics data are a good example, and that is what our experiments use (RNA-seq with methylation, and scRNA-seq with scATAC-seq). Entries are lost to failed acquisitions or dropout, and the simplest fix is to fill them with zeros. The design is reduced to principal components and whitened before PLS, so its columns are orthogonal.

The main message is that it matters which side the missing entries are on. A stronger signal makes up for missing response entries, and in our finite-signal theorem losing half of them costs the same as a signal that is 2\sqrt2 times weaker. It does not make up for missing design entries. With a fraction ρx\rho_x of the design entries observed and αx\alpha_x samples per feature, the squared overlap with the true direction approaches κ=αxρx/(αxρx+1−ρx)<1\kappa=\alpha_x\rho_x/(\alpha_x\rho_x+1-\rho_x)<1 under equal leverage as the signal grows. With two samples per feature, we need more than 60% of the design entries to get within 30∘30^\circ of the true direction. More samples per feature do help, since κ\kappa grows with αx\alpha_x. In general the ceiling depends on the direction, and it can be estimated from the incomplete data (see our response to Reviewer B4ff, Q2).

Q1. Verifying the regimes on real-world data

Only in part, since the experiments use real designs with a synthetic response and an artificial mask, so that the true direction is known. On the TCGA and PBMC designs the direction-specific ceiling predicts high-signal recovery within 0.02 on average (Section 4, Appendix J.5).

A real response with an artificial mask. We tested this on TCGA with methylation as the response, and the result is inconclusive. Without a ground truth we compared against the complete-data direction, which is itself unstable (squared overlap 0.17 between two halves of the samples). With the response fully observed, recovery is about 0.01 below the predicted ceiling when 90% of the design entries are observed and 0.12 below when 30% are. Of the two leading directions, the one with the higher predicted ceiling is recovered better, but it is also the stronger component in the data, so this does not isolate the effect. The real noise may also be correlated, a case where the predictions fail (Appendix K.2). We therefore report the attempt here and not in the PDF.

A real response with a real mask. We cannot test this at present. The true direction is unknown, and the theory covers only entries missing completely at random. Appendix M lists this as untested.

If the main text is still hard to follow, or if there is a further test on real data you would like to see, please let us know so we can address this during the discussion. We hope the revisions address your concerns and that you will consider raising your score.

Dear Reviewer TTS6, thank you again for your review. The discussion closes tomorrow, and we just wanted to check whether our response and the revised submission addressed your questions about practical applications and real data. If anything is still unclear, we'd be happy to clarify!

Reviewer_TTS6OpenReview ↗

Thank you

Thanks for your response. I expected my points to be easily addressed and my original rating reflected that expectation.

We're glad to hear that we've addressed your concerns. Thank you for helping us improve the submission!

Reviewer_B4ffOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies PLS-SVD when missing entries are filled with zeros. It assumes a rank-one regression model, independent missingness in both views, and a complete predictor matrix with orthogonal columns. Dimensions grow proportionally with sample size. As signal strength increases, the response direction is recovered consistently, but the predictor direction can retain a nonzero error. The paper gives an explicit limit for this error and, under additional assumptions, the signal threshold and recovery curves. Synthetic and semi-synthetic experiments test these predictions.

Critical strengths

A precise characterization of masking effects for the PLS estimator. The paper gives precise recovery formulas for this estimator. The high-signal limit depends on how signal energy is distributed across rows with different leverage. This explains why two signal directions in the same design can have different recovery errors. The finite-signal results also quantify the recovery threshold and the behavior above it. These predictions are supported by simulations within the stated model.

Critical weaknesses
  1. The main qualitative finding has limited novelty beyond existing missing data and PLS theory. Missing-covariate regression theory already includes error bounds that depend on coefficient size (Loh and Wainwright, 2012). This literature is particularly relevant because the paper studies a regression model with corrupted predictors and additive response noise. Here, increasing signal strength amplifies both the useful signal and the fluctuations caused by missing predictors. Persistent angular error is therefore natural. PLS inconsistency in proportional dimensions is also established (Chun and Keleş, 2010, Theorem 1). The paper needs to explain more clearly what those formulas add beyond the established qualitative behavior. From the current presentation, the additional statistical insight appears limited.

  2. The presentation makes the contribution unnecessarily hard to assess. The manuscript runs to over 70 pages, yet the main qualitative phenomenon can be explained much more directly using the existing missing-covariate and high-dimensional PLS literature. For example, Mergny and Zdeborová (2026) already establish PLS recovery thresholds and overlaps. Section 2 briefly summarizes this work, but does not explain concretely what the new results add. The comparison should distinguish changes caused by missingness from differences between the earlier two-noisy-view model and the present regression model with a noise-free complete design. It should then explain which statistical conclusions are new and why they matter. Without this comparison, the reader must reconstruct the contribution from separate papers. This is particularly burdensome in a manuscript of over 70 pages, whose central message could be presented much more directly.

  3. The qualitative asymmetry is largely due to the authors' regression model. PLS-SVD treats the two views symmetrically, but the model pairs a full-rank, whitened predictor matrix with a rank-one response signal. Even without additive noise, masking the predictors can rotate the recovered predictor direction while leaving the response direction exact when the response is fully observed. This gives a simple explanation for the basic asymmetry; it should be interpreted within this model instead of as a general property of PLS. The scope also matters: the complete design requires Dx≤ND_x \le N, and the masked finite-signal theorem’s additional spectral assumption is verified for nested flat orthogonal designs such as Hadamard constructions.

  4. The experiments establish accuracy within the model more clearly than statistical usefulness beyond it. In the biological examples, complete data are reduced and whitened before synthetic responses and missing entries are generated. These are useful tests of design geometry, but they do not evaluate recovery from incomplete raw data or prediction of actual outcomes. Appendix I.5 includes imputation and prewhitening comparisons, although the theory does not explain their relative performance or help select among them. The evidence mainly confirms the formulas for zero-filled PLS-SVD under the assumed model, leaving their wider practical benefit unclear.

Questions for authors
  1. Please compare Theorems 3.3–3.4 with the recovery results of Mergny and Zdeborová (2026). Which conclusions are already established in the earlier setting, what changes because of masking, and what differs because the underlying model has changed?

  2. Can the limiting error be estimated from the incomplete data, without knowing the complete predictor matrix or the true signal direction? Or is the formula intended only as a reference when those quantities are known?

5 comments

Thank you for your careful and extensive review. Your detailed review also merits a longer response, which we hope is okay. The primary issue you raised was that the paper was difficult to follow, which we have worked to address. This can be seen in the revised submission (changes in blue). We have also added further comparisons to other work, as you requested, in Section 2, explained the ceiling directly in the Introduction, and moved several technical passages out of Section 3.

W1. What the formulas add beyond known missing-data and PLS results

We agree that a persistent angular error is natural, and we now say so in the Introduction. What the theorems add is the exact value of the limit (the ceiling), its dependence on the direction, and the behaviour at finite signal (see answwe to Q1).

Section 2 now discusses both papers you cite. Loh and Wainwright (2012) bound the error of a sparse estimator that corrects for the mask. As you note, the bound grows with the coefficient norm, but it vanishes when klog⁡p/n→0k\log p/n\to0. Our direction is dense, the number of features is proportional to the number of samples, and the estimator is uncorrected. The error therefore does not vanish, and Theorem 3.3 gives its exact high-signal limit.

Chun and Keles (2010) show that the PLS regression coefficient is inconsistent when p/np/n does not vanish, with complete data and without giving a limit. In our model with complete data, the overlap with the true direction follows a classical formula and tends to 1 as the signal grows (Corollary 3.5). With a masked design it stays below 1, which we call that limit the ceiling.

Q1. What is known, what the model changes, and what masking changes

Known, with complete data. Mergny and Zdeborova (2026) study two views with independent Gaussian noise in each and correlated low-rank spikes, and give the detection threshold and the limiting overlaps. Leger and Chatelain (2025) treat a model with shared and view-specific components. In such models each view's overlap is limited by its own signal-to-noise ratio, so the design-side overlap stays below 1 when only the response-side signal grows (the diagonal case of Leger and Chatelain). Our ceiling has the same form which we now state in Section 2. Neither of the papers treat missing entries.

Changed by the model. Our design is noise-free and column-orthogonal. With complete data both overlaps then tend to 1 as the signal grows (Corollary 3.5, the classical result of Benaych-Georges and Nadakuditi, 2012), so without the mask there is no ceiling. The noise-free design is also what lets the proofs condition on the mask. With noise in both views the conditioning fails, so our results do not carry over to a model with two noisy views.

Changed by masking, at high signal (Theorem 3.3). The ceiling depends on the direction. The mask acts as noise on the design, and the resulting signal-to-noise ratio is weighted by the leverage of the samples that carry the signal. Two directions in the same design therefore have different ceilings. On the biological designs the direction-specific ceiling predicts high-signal recovery within 0.02 on average. Without the weighting the average error is between 0.05 and 0.08.

Changed by masking, at finite signal (Theorem 3.4). The threshold and the recovery curves are the main technical contribution. In the two-view models the design is a low-rank spike in isotropic noise. Here the masked design is a full-rank matrix plus noise. Both the signal direction and the noise in X⊤YX^\top Y depend on the design mask. The standard theory of low-rank perturbations assumes that the perturbation is independent of the noise, so it does not apply.

The threshold also differs from the one obtained by treating masking as a loss of signal-to-noise ratio alone, as one does for PCA. We call this prediction the attenuation-only reference, and it is the gray dotted curve in Figure 1a and b. In Figure 1(a) it puts the threshold at 1.67, and ours is 1.39. In Figure 1c, on random orthogonal designs, the measurements agree with our prediction and exclude the reference in every tested case with at most 45% of the design entries observed. Appendix J.3 reports the other cases. When 70% or more are observed, the measurements lie 1 to 2% below our prediction.

New statistical conclusions. A stronger signal makes up for missing response entries but not for missing design entries. The loss from missing design entries depends on the direction, and more samples per feature reduce it where a stronger signal cannot.

W2. The comparison with Mergny and Zdeborova and the length of the manuscript

The comparison of the results is under Q1 in our first post and is now in Section 2.

Differences from their setting. Mergny and Zdeborova have a low-rank signal in each view and Gaussian noise that is independent of the signal, and our model has neither once the design is masked. The masked design is a full-rank matrix plus noise from the mask. The same mask sets the signal direction and the noise covariance of X⊤YX^\top Y, so the signal and the noise are dependent.

Why the proof is long. We handle the dependence by conditioning on the design mask. The noise covariance then depends on the design and the mask, and three steps that are standard for isotropic Gaussian noise have to be proved. First, we need the spectral limits of this covariance along the signal direction (Appendices E and G). Second, the masked response noise is not Gaussian, so we need a comparison with Gaussian noise. Third, we must show that no noise singular value leaves the bulk, so that the outlier can be attributed to the signal (Appendix F). We agree that the proof is long and have included the details for clarity and completeness. The heuristic replica derivation provides a shorter route to the same equations in seven pages (Appendix H).

W3. The asymmetry comes from the regression model, and the scope of the theorems

Yes, the asymmetry comes from the regression model and is not a general property of PLS. PLS-SVD treats the two views symmetrically, but in our model only the response is noisy, and the Introduction states that the asymmetry is specific to this model.

Your noise-free argument explains the rotation, and we give the size of the rotation in Proposition C.3. Theorem 3.3 shows that it has the same size in the high-signal limit when the response is noisy and masked. The argument does not give the threshold or the recovery curves at finite signal, which are the content of Theorem 3.4.

The symmetric model with noise in both views is harder because one latent score generates both views. The noisy design, the recovered direction and the noise in X⊤YX^\top Y are then coupled, and conditioning on the design mask no longer fixes the signal direction and the noise covariance, which is the step our proofs rest on. We list this as an open problem in Appendix M and see it as the natural next step, for us or for others.

We also agree on the scope. The complete design needs at least as many samples as features, because its columns are orthogonal, and we have no results beyond that. The spectral condition of Theorem 3.4 is proved for designs taken from nested Hadamard matrices. For other designs with missing design entries the curves are predictions, and in simulation they hold on random orthogonal designs (Figure 1c, Appendix J.3) and on the biological designs at high signal. Table 1 in Appendix A lists what is proved for which class of designs.

W4. The experiments test the formulas within the model, not their practical use

Yes, the experiments test the formulas within the model and not their practical use, and we say so in Appendix M. The biological experiments use a real design with a synthetic response so that the true direction is known.

On incomplete raw data, the revision adds a test in which a raw design with correlated columns is masked and then whitened (Appendix L.3). The ceiling drops by up to 0.18, and an entry-wise version of our formula still predicts it within 0.003. We also tried a real response from TCGA, but without a ground truth the result is inconclusive (see our response to Reviewer TTS6, Q1).

We do not claim that zero-filling is the best estimator, and the theory does not rank the alternatives in Appendix K.3 (I.5 in the submission). What it does give is the ceiling of zero-filling, which can now be estimated from the incomplete data (Q2) and shows when a stronger signal will not help.

Q2. Estimating the ceiling from the incomplete data alone

Yes, within the model. The revision adds an estimate that uses only the masked data (Section 4, Appendix L.2, Table 6). It estimates each sample's leverage from the observed entries and its score from the response, and it needs neither the complete design nor the true direction. At high signal its mean over trials is within 0.005 of the true ceiling on the engineered designs and on the eight leading TCGA directions.

The estimate gives the high-signal limit and not the error at a given signal strength. It also assumes homogeneous masking and a design whitened before masking, and the whitening cannot be checked from the incomplete data.

If any parts of the paper remain difficult to follow, or if you still believe that a conclusion of Theorem 3.3 or 3.4 is covered by earlier work, please let us know so we can address this during the discussion. We hope these changes address your concerns and that you will consider raising your score.

Thank you for your response. However, I would like to maintain the score given the model settings is simplified as orthogonal design, rank 1 model and additive noise, etc, which is not a full theory for two view PLS. I also feel the revised 79 pages are hard to read.

Comments on the paper

1 comment

Thank you all for taking the time to read our paper and discuss it with us. Your comments helped us see what needed clearer explanations and where further experiments would be useful. We've revised the submission with changes marked in blue. Here is a brief summary of the main changes:

  • Main contributions. We've worked to explain more clearly what our results add: The exact recovery ceiling for each signal direction, along with the recovery threshold and curves at finite signal strength. The new calculations and simulations in Appendix L also explore how far these predictions extend beyond the assumptions of our theorems and showing where they continue to hold and where they break down.

  • Comparison with earlier work. We've expanded Section 2 to explain how our results relate to those of Mergny and Zdeborova and Leger and Chatelain. More precisely, we distinguish the effects of masking from differences between the underlying models. Figure 1a also shows why treating masking simply as a loss of SNR does not capture the full recovery curve.

  • Clearer presentation. We've revised the "Introduction" and "Conclusion" sections to explain the setting, results, and practical implications more directly. We've also shortened the technical discussion in Section 3, simplified the explanation of the spectral condition and proof outline, and reorganized the experimental appendices so the setup, validation, robustness tests, and extensions are easier to find.

  • Practical interpretation and estimating the ceiling. In response to the question of whether the ceiling can be estimated from incomplete data, we've added an estimator in Section 4 and Appendix L.2. Within our model, it requires neither the complete design nor the true signal direction and we test it on both engineered and biological designs. We've also added more discussion of what the ceiling means in practice and when stronger signals or more samples can help.

  • Extensions and limitations. Appendix L now includes calculations and simulations with multiple components, different missingness patterns and whitening after masking. We discuss which results appear to carry over and where the predictions break down. These calculations are separate from the theorems and estimating the whitening from incomplete data and testing on naturally incomplete data remain open problems, as discussed in Appendix M.

  • Correction to the Hadamard experiments. Thank you, Reviewer ewDp, for catching an error in how the Hadamard columns were selected. We've corrected it and rerun Figure 1a and Figure 2 using the nested design covered by our theory. The conclusions remain unchanged and no theorem or proof was affected.

We believe these revisions address the main concerns raised and make our contributions and limitations clearer. Regarding a theorem for several components, Appendix L shows that parts of our analysis and proof already extend beyond rank one. A full recovery theorem at finite signal strength would require new arguments to handle the interactions between components, which would be a significant piece of work on its own. We see our results as a concrete starting point for this further work. If anything remains unclear, we'd be happy to clarify it before the discussion closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.