Blind Noise Equalization for Large Data Matrices under General Variance Profiles
Authors: Boris Landa, Francesco Strino, Fabio Parisi, Ruiqi Li, Yuval Kluger
Organizations: Department of Electrical and Computer Engineering, Yale University · Program in Applied Mathematics, Yale University · PCMGF Limited, United Kingdom · Computational Biology and Biomedical Informatics Program, Yale University · Department of Pathology, Yale School of Medicine
Low-rank approximation is a basic tool for inference from large and noisy data matrices. However, when noise variances differ across entries, a setting known as heteroskedasticity, the noise can distort the observed spectrum, producing spurious components and obscuring weak signals. To address these challenges, we develop BRINE (Blind Resolvent-based Iterative Noise Equalizer), a data-driven procedure that iteratively rescales rows and columns to equalize the average noise variance across both dimensions. Under suitable conditions, this procedure restores the Marchenko-Pastur law for the noise spectrum and can amplify signal components without prior knowledge of the signal or noise structure. A key idea underlying our approach is to equalize the diagonal entries of the data resolvent as a proxy for equalizing the unknown noise variances. At the population level, we identify a potential function that characterizes this equalization and motivates an iterative minimization procedure. We prove convergence of the iterations to the true equalizer for arbitrary positive variance profiles, with a geometric rate under mild conditions. The updates can be estimated directly from noisy data, and we establish probabilistic error bounds for the scaling factors. Simulations and experiments across diverse biological data modalities and preprocessing pipelines illustrate the effectiveness of our approach.
Figures & tables
Figure 1: Stylized example with m=1000 , n=2000 , and a rank- 20 signal containing ten strong and ten weak components. The noise is Gaussian with the variance profile shown in Panel A: a 2×2 block structure with ten scattered high-variance rows and ten scattered high-variance columns (indicated by triangles). Panel B shows the leading eigenvalues of YYT . The blue bars are the observed eigenvalues, while the filled and open black markers show the eigenvalues of XXT associated with the strong and weak signal components, respectively. The horizontal lines mark the MP upper edge (dashed purple), the upper edge of the main noise bulk (dashed red), and the largest eigenvalue of the noise matrix EET (dotted green). The vertical line marks the true rank. Panel C compares the empirical spectral distribution of the noise with the MP law. Panels D–F illustrate the benefits of our proposed method. Panel D shows the estimated potential and the Kolmogorov-Smirnov (KS) statistic for goodness-of-fit to the MP law over the iterations. Panel E shows the leading eigenvalues after applying the proposed normalization, with the MP upper edge shown by the horizontal line. Panel F compares the empirical spectral distribution of the normalized noise with the MP law.
Figure 2: Scaling-factor estimation accuracy in the same signal-plus-noise setup as for Figure 1 . The vertical axis in both panels is the logarithmic error d , defined in ( 27 ), between the estimated factors and the true scaling factors (x⋆,y⋆) . Panel A. Error versus iteration for m=4000 , n=8000 , obtained from the data-driven procedure (BRINE) and the population-level procedure (computed via the true variance profile), together with the errors of the rank-one Dyson equalizer [ 42 ] and the empirical-variance equalizer [ 44 ] (which are independent of iterations). The trajectories are displayed through iteration 25 . Panel B. Error versus the number of rows m , with n=2m , for BRINE with maximum iteration caps T=10 and 25 , and the same two alternative methods as in Panel A.
Figure 3: Effect of BRINE normalization on the empirical spectral distribution across four representative biological modalities. Top row: before normalization. Bottom row: after normalization. Each panel overlays the scaled empirical spectral distribution (histograms) with the theoretical MP density (red curve); green dashed lines mark the MP upper edge. Different histogram colors correspond to different preprocessing transformations applied to the raw data. Complete results for all 11 datasets are provided in Figure 7 and Figure 8 in the appendix.
Figure 4: Simulated departures from the assumptions of independent noise and low-rank signal structure. Panel A. The effect of correlated noise on the MP fit, for m=n=1000 and a rank- 10 signal. The heatmap shows the Kolmogorov–Smirnov statistic over a 20×20 grid of the mixing parameters in ( 48 ), with the expected proportion f of nonzero off-diagonal entries on the vertical axis and their intensity h on the horizontal axis. The bottom-left corner corresponds to i.i.d. noise. Panel B. The largest eigenvalue of EET , using scaling factors estimated from Y , as a function of signal rank for several signal-to-noise ratios (SNRs). Here, m=5000 and n=10000 , and the dashed line marks the MP upper edge β+ . Curves show the mean over five realizations, with error bars extending two standard deviations on either side of the mean.
Figure 5: Recovery of a rare cell population as a function of the target size s for the five preprocessing pipelines. Left to right: precision, recall, and specificity, as defined in ( 49 ). Each panel plots one minus the metric, so lower values indicate better recovery. Both axes are logarithmic, and LOWESS curves show the trends.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Comparison of BRINE to [ 42 ] and [ 44 ] under the simulation setup of Figure 1 . Each panel shows the empirical spectral distribution of the noise after applying the scaling factors estimated from the observed signal-plus-noise matrix, together with the MP density and its upper edge. A. BRINE closely restores the MP law, with KS distance 0.02 and no noise eigenvalues above the MP edge. B. The rank-one Dyson equalizer of [ 42 ] , derived under a rank-one variance-profile model, leaves a visible upper tail, with KS distance 0.100 and 93 noise eigenvalues above the MP edge. C. The empirical row/column variance equalization procedure of [ 44 ] produces no noise eigenvalues above the MP edge but substantially compresses the noise spectrum due to the presence of strong signals in the data, resulting in a KS distance of 0.313 .
Figure 7: Empirical spectral distributions before normalization across all datasets and modalities. Each panel corresponds to one dataset; different histogram colors represent different preprocessing transformations. The red curve indicates the theoretical Marchenko–Pastur (MP) density, and green dashed lines mark the MP upper edge β+ . Spectra are scaled by a constant for visualization.
Figure 8: Empirical spectral distributions after normalization for the same datasets and transformations as in Figure 7 . After normalization, all spectra closely follow the theoretical MP distribution (red curve) regardless of modality or preprocessing. Green dashed lines mark the MP upper edge β+ .
Figure 9: Convergence of BRINE normalization across all datasets. Each row corresponds to one dataset; each color represents a distinct preprocessing transformation. The three columns show the Kolmogorov–Smirnov p -value, KS statistic, and the estimated potential J^t over successive iterations.
Figure 10: Convergence of BRINE normalization across all datasets (cont).
Figure 11: Convergence of BRINE normalization across all datasets (cont).
Spectral methods rely fundamentally on the stability of principal eigenspaces under random perturbations. Classically, this stability is quantified by the Davis-Kahan and Wedin theorems, which bound the eigenspace error using the operator norm of the noise and the relevant spectral gaps. While these worst-case bounds are sharp for arbitrary deterministic perturbations, they can be wasteful in the low-rank signal-plus-random-noise setting, as they fail to capture the fine-grained interaction between the signal geometry and the noise distribution. In this paper, we study the spectral perturbation of signal-plus-noise matrices corrupted by sparse, random noise with an arbitrary, inhomogeneous variance profile. We demonstrate that under heterogeneous noise variances, the empirical eigenvectors suffer a systematic, deterministic geometric bias that is entirely invisible to classical perturbation bounds. By leveraging the Quadratic Vector Equation (QVE) and establishing fine-grained isotropic local laws, we derive near-optimal, non-asymptotic perturbation bounds for the leading eigenspaces in the operator and 2→∞ norms. The bounds separate the usual signal-to-noise contribution, stochastic fluctuations, and structured geometric bias terms determined by the alignment between the signal eigenspaces and the row-wise variance profile.
Fengkai Liu, Ke Wang, Wanjie Wang
Department of Mathematics, Hong Kong University of Science and Technology · Department of Statistics and Data Science, National University of Singapore
We study transfer learning for low-rank plus sparse matrix denoising when both ambient and latent dimensions grow. A data-rich source task provides subspace information for a higher-dimensional, data-scarce target with a small rank increment and localized sparse edits. We develop an anchored alternating-projection estimator that preserves transferred subspaces while learning target-specific coefficients, additional latent directions, and sparse modifications. We establish deterministic error bounds that separate target noise, representation growth, and source-estimation error, identifying improved rates through small structural increments and reduced noise within inherited subspaces. We further extend it to approximate anchoring quantifies robustness to subspace mismatch. We establish application-specific guarantees for Markov transition matrix estimation from single trajectories with dependent noise and for structured covariance estimation, with the latter analysis provided in the appendix. Experiments demonstrate consistent improvements over non-transfer baselines on synthetic data and an approximately 45% reduction in principal-angle error on a NYC taxi transition-estimation task with expanding state spaces.
Elynn Chen, Jinhang Chai, Xuyuan Liu +1
Department of Technology, Operations, & Statistics, New York University · Department of Computer Science, Dartmouth College
Equalizer parameter optimization is critical for signal integrity in high-speed memory systems operating at multi-gigabit data rates. However, existing methods suffer from computationally expensive eye diagram evaluation, optimization of expected rather than worst-case performance, and absence of uncertainty quantification for deployment decisions. In this paper, we propose a distributional risk-sensitive reinforcement learning framework integrating Information Bottleneck latent representations with Conditional Value-at-Risk optimization. We introduce rate-distortion optimal signal compression achieving 51 times speedup over eye diagrams while quantifying epistemic uncertainty through Monte Carlo dropout. Distributional reinforcement learning with quantile regression enables explicit worst-case optimization, while PAC-Bayesian regularization certifies generalization bounds. Experimental validation on 2.4 million waveforms from eight memory units demonstrated mean improvements of 37.1% and 41.5% for 4-tap and 8-tap equalizer configurations with worst-case guarantees of 33.8% and 38.2%, representing 80.7% and 89.1% improvements over Q-learning baselines. The framework achieved 62.5% high-reliability classification eliminating manual validation for most configurations. These results suggest the proposed framework provides a practical solution for production-scale equalizer optimization with certified worst-case guarantees.
Muhammad Usama, Dong Eui Chang
Control Laboratory, School of Electrical Engineering, KAIST, Daejeon, 34141, Republic of Korea