Pith. sign in

REVIEW 3 major objections 5 minor 46 references

A Bayesian Framework for Regularized Estimation in Multivariate Models Integrating Approximate Computing Concepts

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Quantization error can be recast as Bayesian prior uncertainty

desk verdict The paper's central 'quantization–Bayes' bridge is a Gaussian additivity tautology, but the regularized LDA algorithms and the experimental comparisons are a solid kernel that could survive a major rewrite. read the letter →

arxiv 2509.10045 v1 pith:GU4KUAP2 submitted 2025-09-12 stat.ME

Jan Kalina This is my paper · ORCID
classification stat.ME MSC 62F1562H3062J07
keywords Bayesianestimationquantizationshrinkageestimatorsregularizedlineardiscriminantanalysismultivariatenormalapproximatecomputingconjugatepriorsposteriorequivalence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish a formal bridge between two normally separate ideas: rounding parameters to finite precision (quantization) and Bayesian estimation. It claims that adding a Gaussian noise term to represent rounding is mathematically equivalent to inflating the variance of a prior distribution, so that a quantized estimator is literally the same object as a Bayesian posterior estimator. If true, this means the uncertainty introduced by approximate computing—post-training rounding, low-precision arithmetic—can be handled with standard Bayesian tools, without any new methodology. The same Bayesian viewpoint is then used to justify shrinkage estimators of group means in linear discriminant analysis, and the paper proposes efficient algorithms for these regularized classifiers. A sympathetic reader would care because the result offers a unified probabilistic language for approximation error and regularization, potentially guiding the design of low-precision but statistically principled models.

What carries the argument

The load-bearing identity is Gaussian additivity under independent increments: if μ ∼ N(θ, Ψ) and τ ∼ N(0, δ²I) are independent, then ξ = μ + τ ∼ N(θ, Ψ + δ²I). This single implication carries the entire quantization–Bayes bridge, because it makes the posterior of ξ under the quantized model coincide with the posterior under a model with an inflated prior covariance. The companion machinery is conjugate normal–normal Bayesian updating, which yields closed-form shrinkage estimators of the form (I − Δ)X̄ + Δθ; the shrinkage matrix Δ is a ratio of the prior and data precisions. These two pieces together turn quantization, shrinkage, and regularized discriminant analysis into instances of the sa

What would settle it

Run a simulation with n i.i.d. p-dimensional observations, compute the MLE, apply deterministic rounding at a fixed bit width (e.g., 8 bits), and compare the resulting estimator against the posterior estimator of the Gaussian-perturbation model with matching δ²; if the two diverge substantially as the bit width shrinks, the claimed equivalence does not describe actual quantization.

Watch

Extended reading notes

Core claim

The central claim of the paper is a formal equivalence between two seemingly different models. In Model 3, the true parameter μ is random with prior N(θ, Ψ) and the estimator is perturbed by an independent quantization error τ ∼ N(0, δ²I); in Model 4, the parameter is directly assigned the prior N(θ, Ψ + δ²I) and no explicit quantization is introduced. Theorem 2 proves that the posterior estimator of the parameter is identical in both models; Theorem 1 establishes the corresponding result in the non-Bayesian case where μ is fixed. This means quantization can be reinterpreted as a particular form of prior variance inflation, turning a deterministic technical artifact into a probabilistic mode

Load-bearing premise

Rounding to finite precision can be represented as an independent additive Gaussian perturbation added to the estimated parameter.

Editorial extensions

If this is right

  • If quantization is accepted as Gaussian prior inflation, then any estimator that rounds parameters to fixed precision can be analyzed, and its uncertainty propagated, using standard Bayesian posterior formulas.
  • Shrunken group means in linear discriminant analysis are exactly posterior means under normal priors centered at a pooled mean, so their regularization gains a rigorous probabilistic justification.
  • The proposed regularized LDA variants—ridge, soft-thresholding, and hard-thresholding means—are computable without explicit matrix inversion via Cholesky or SVD, at O(p³) cost, and the results hold for finite samples.
  • Numerical experiments indicate that hard-thresholded means with a non-diagonal shrinkage target give the best classification accuracy and the most parsimonious variable selection on both the real and simulated data.
  • The equivalence implies that tuning the quantization precision and tuning the prior covariance are alternative dials for the same regularization effect, which could simplify model selection.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The Gaussian representation of rounding is an idealization: actual rounding is deterministic and discrete, so the bridge holds only insofar as a Gaussian perturbation is a faithful proxy; replacing it with a uniform rounding-error model would yield a different posterior, not a simple variance inflation.
  • The link suggests a practical recipe for deep-learning post-training quantization: choose the bit width so that the implied prior variance matches the regularization level that minimizes a Bayesian risk, thus connecting hardware constraints to statistical inference.
  • A testable extension is to simulate data with true deterministic rounding at decreasing precision and compare the posterior mean under the Gaussian-perturbation model; a divergence at low bit widths would bound the practical range of the equivalence.
  • Because the best-performing covariance target in the experiments matched the simulation's generating design, the empirical advantage of that target may partly reflect favorable design alignment rather than a universal property; this is consistent with the author's own observation that target choice is consequential.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a Bayesian framework for regularized estimation in multivariate models, with three advertised components: (i) a link between Bayesian estimation and parameter quantization/rounding, (ii) a Bayesian justification of shrinkage estimators for two-group means, and (iii) regularized linear discriminant analysis (LDA) algorithms with mean and covariance shrinkage. The quantization link is formalized in Section 3 through Models 1–4, where quantization error is represented by an independent Gaussian perturbation τ ∼ N(0, δ²I) added to the parameter. Theorems 1 and 2 claim that the posterior estimator of the perturbed parameter ξ in these models coincides with that in models with a Gaussian prior on ξ of inflated variance. Sections 4–5 derive shrinkage estimators for group means and propose Algorithms 1–2 for regularized LDA. Section 6 reports simulation and real-data experiments comparing LDA variants with different shrinkage targets and mean regularizations.

Significance. If the quantization–Bayes connection were valid, the paper would offer a bridge between approximate computing and Bayesian inference, potentially justifying Bayesian treatments of quantized parameters. The paper also includes correct conjugate-posterior derivations (Lemmas 1–2), clearly stated LDA algorithms, and reproducible-looking experimental results. However, the central novelty is not established: the quantization link is only a trivial Gaussian variance-inflation identity, because no rounding operator or quantization error model actually appears in the formal setup. The remaining contributions—Bayesian shrinkage for means and regularized LDA—are standard and do not depend on the quantization claim. The experimental comparison is further weakened by a simulation design that uses the same covariance structure as the preferred shrinkage target T2. On balance, the advertised conceptual contribution is unsupported, and the residual contributions are insufficient for publication in a serious statistics journal.

major comments (3)
  1. [Section 3, Eqs. (10), (13); Theorems 1–2] The central claim that quantization is connected to Bayesian estimation is unsupported. Quantization (rounding) is a deterministic, discrete, parameter-dependent map Q(·), but the models replace it by an independent additive Gaussian perturbation τ ∼ N(0, δ²I). No quantizer, discrete alphabet, or rounding-error bound appears anywhere in Models 1–4. Theorem 1 and Theorem 2 are true only as statements of Gaussian additivity: if ξ = μ + τ with τ independent Gaussian, then ξ is Gaussian with variance Ψ + δ²I. Indeed, Model 3 defines the observation equation as Xi = ξ + ε_i and ξ = μ + τ with μ ∼ N(θ,Ψ), which is identical by construction to Model 4's prior ξ ∼ N(θ,Ψ+δ²I) and likelihood Xi = ξ + e_i. The theorems therefore provide no bridge to approximate computing. The paper's own limitation paragraph at the end of Section 3 concedes that rounding errors could alternatively be modeled as uni
  2. [Section 6, Table 1] The simulation study is biased in favor of the shrinkage target T2. The data are generated from Σ = σ²[(1−c)I_p + c 1_p1_pᵀ] with c = 0.4, while T2 = σ²I_p + ϑ(1_p1_pᵀ − I_p) with ϑ = 0.15. Both matrices have the same compound-symmetric form (a linear combination of I_p and J_p), so the target T2 lies in the same family as the true covariance. The statement that T2 'precisely corresponds to the data-generating design' is numerically inaccurate (T2 ≠ Σ unless ϑ = c, which is not the case), but even the weaker claim of 'corresponding structure' means the comparison between T1 and T2 is not a fair test of the methods. The conclusion that T2-based LDA is superior is therefore at least partly an artifact of the simulation design.
  3. [Lemma 2, Eq. (24)] Lemma 2, the multivariate extension of the shrinkage result, is stated without proof. Since this result is used to justify the regularized means in Section 5, it is load-bearing for the LDA contribution. The proof is a straightforward matrix generalization of Lemma 1, but it should be supplied or a precise reference given. As written, the manuscript asserts a key result without demonstration, which is especially conspicuous given that the proof of Lemma 1 contains a cross-reference error (pointing to Eq. (24) instead of Eq. (17)).
minor comments (5)
  1. [Section 4, proof of Lemma 1] The sentence 'which can be equivalently reformulated as (24)' should refer to Eq. (17) in the univariate Lemma 1, not Eq. (24) of Lemma 2.
  2. [Definition 1, Eq. (31)] The hard-thresholding rule has an inconsistency at the boundary: the indicator uses |X_k| > δ(0), while the piecewise expression assigns the value X_k when |X_k| ≥ δ(0) and 0 when |X_k| ≤ δ(0). The equality case should be specified consistently.
  3. [Section 6, Table 1] The reported 'standard deviation' is computed from 5-fold cross-validation, i.e., from only five accuracy numbers. This is not a meaningful measure of variability of the classification error, and it does not support comparative statements such as 'overcome those of [2] remarkably' without a statistical test or confidence intervals.
  4. [Section 3, after Theorem 2] The phrase 'paradigm change' and 'significant shift in perspective' is overclaiming, even if the quantization link were valid; the paper itself acknowledges several limitations that directly qualify this claim.
  5. [References] Reference [26] is cited as a working paper, while a published version [27] exists; the experimental use of 'Ledoit–Wolf asymptotics' should refer to the published source.

Circularity Check

2 steps flagged · score 8.0 of 10

Section 3's quantization–Bayes bridge is a Gaussian-additivity tautology: Models 1–4 define quantization error as an independent Gaussian perturbation, so Theorems 1–2 restate the convolution of independent Gaussians by construction.

  1. self definitional [Section 3, Model 1 (Eq. 10) and Theorem 1]
    "Instead of the true model (9), the parameter estimation will be rather considered in the model Xi = ξ + ε_i, i = 1, ..., n, where ξ is obtained as ξ = μ + τ (10) with a random vector τ ∼ Np(0, δ^2 I_p) independent of ε_1, ..., ε_n. ... Theorem 1. The posterior estimator of ξ in Model 1 is the same as the posterior estimator of ξ in Model 2. Proof. The idea is that a random vector ξ defined as ξ=μ+τ, where μ is a constant vector and τ∼N(0, δ^2I_p), satisfies ξ∼N_p(μ, δ^2I_p)."

    Quantization/rounding is never represented by a rounding map or discrete alphabet; it is replaced by an independent Gaussian perturbation τ. Model 2 defines ξ ∼ N(μ, δ²I), which is exactly the distribution of μ + τ. Hence the theorem merely restates the definition of Gaussian additivity. It cannot establish a connection between rounding/quantization and Bayesian estimation, because the quantizer itself never appears.

  2. self definitional [Section 3, Models 3–4 (Eqs. 13–14) and Theorem 2]
    "Model 3. ... Xi = ξ + ε_i, i = 1, ..., n, where ξ is obtained as ξ = μ + τ (13) with a random vector τ ∼ Np(0, δ^2 I_p) independent of μ and ε_1, ..., ε_n. Model 4. ... ξ ∼ Np(θ, Ψ + δ^2 I_p) ... Theorem 2. The posterior estimator of ξ in Model 3 is the same as the posterior estimator of ξ in Model 4. Proof. ... ξ defined as ξ=μ+τ fulfils ξ∼Np(θ,Ψ+δ^2I_p)."

    Model 3 defines ξ as μ + τ with μ ∼ N(θ,Ψ) and τ ∼ N(0,δ²I) independent; by elementary convolution ξ ∼ N(θ, Ψ + δ²I), which is precisely Model 4. The equality of posterior estimators is therefore true by construction, not a derived bridge between rounding and Bayesian inference. The paper later concedes the stochastic Gaussian model is one possible modeling choice ('alternatively, rounding errors could be modeled stochastically'), underscoring that the theorem cannot follow from the actual deterministic rounding operation.

full rationale

The paper's advertised novelty is the claimed connection between quantization (approximate computing) and Bayesian estimation, stated in the Abstract and Section 3 and formalized in Theorems 1–2. That connection reduces, by the paper's own definitions, to a one-line Gaussian convolution identity. In Model 1 (Eq. 10) and Model 3 (Eq. 13), 'quantization' is modeled as ξ = μ + τ with τ ∼ N(0, δ²I) independent of other terms. Models 2 and 4 then simply declare ξ to have the resulting normal distribution (N(μ, δ²I) or N(θ, Ψ+δ²I)). Theorems 1 and 2 prove the equality of posterior estimators under these two equivalent specifications, which is true by definition and does not involve any rounding operator, quantization grid, or deterministic map. Thus the central 'paradigm change' claim is circular: the conclusion (quantization error behaves like Gaussian prior-variance inflation) is assumed in the model setup. The paper itself flags the Gaussian model as uncertain, noting rounding errors could instead be modeled as uniform perturbations. By contrast, the shrinkage-mean derivations (Section 4, Lemmas 1–2) are standard conjugate-posterior calculations and are not circular; the LDA algorithms and experiments (Sections 5–6) are self-contained and do not depend on the Section 3 results. Self-citations appear but are not load-bearing. The overall score is therefore high because the headline claim is forced by definition, even though other parts of the paper are independent.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claims rest on standard Gaussian conjugacy plus an ad hoc assumption that quantization can be represented as independent Gaussian noise. The regularization parameters λ and δ are not derived but selected by cross-validation, and the simulation favors the hand-chosen target T2. The invented Gaussian noise term has no independent empirical support.

free parameters (4)
  • δ (variance of Gaussian quantization noise τ) = not estimated
    Introduced in Models 1 and 3 (Eqs. 10, 13) to model rounding as Gaussian noise; essential to Theorems 1-2 but not fitted to any data.
  • λ (covariance shrinkage parameter) = selected by cross-validation or Ledoit-Wolf
    Used in regularized covariance estimators (8), (35), (39); the paper uses 5-fold CV or asymptotic values.
  • δ^(2), δ^(1), δ^(0) (mean shrinkage/threshold parameters) = selected by cross-validation
    Shrinkage strengths for L2, soft-thresholding, and hard-thresholding in Definition 1, chosen to maximize classification accuracy.
  • ϑ = 0.15 in covariance target T2 = 0.15
    Hand-chosen constant in T2 = σ²I + ϑ(11ᵀ−I); the simulation data are generated with this exact target, biasing comparisons.
assumptions (4)
  • standard math Gaussian conjugacy of the normal prior with the normal likelihood
    Used in Lemmas 1-2 and Eq. (5) to derive closed-form posterior means as shrinkage estimators; a standard textbook result.
  • ad hoc to paper Independent additive Gaussian perturbation represents quantization error
    Central to Theorems 1-2; actual rounding is deterministic/discrete and the paper offers no evidence that Gaussian noise approximates it; the paper later hedges this in its limitation statement.
  • standard math Independence of samples and priors across groups
    Assumed in Lemmas 1-2 for the two-sample shrinkage derivation; standard and unobjectionable.
  • domain assumption Common covariance matrix Σ across groups in LDA
    Assumed for all LDA formulations in Section 5; standard LDA assumption, noted by the paper as a limitation when group covariances differ.
invented entities (1)
  • Gaussian quantization noise τ
    purpose: Modeled as N(0, δ²I) added to the parameter ξ = μ + τ to claim that rounding/quantization can be viewed as prior variance inflation (Theorems 1-2).
    No falsifiable prediction connects τ to actual rounding error; it is a purely mathematical device chosen for conjugate convenience, not validated against any measured quantization process.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Bayesian Framework for Regularized Estimation in Multivariate Models Integrating Approximate Computing Concepts." pith.science (2026). https://pith.science/paper/GU4KUAP2

@misc{pith2026250910045,
  author       = {Pith},
  title        = {Pith review of: A Bayesian Framework for Regularized Estimation in Multivariate Models Integrating Approximate Computing Concepts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GU4KUAP2}},
  note         = {Machine review of arXiv:2509.10045}
}
read the original abstract

This paper discusses regularized estimators in the multivariate statistical model as tools naturally arising within a Bayesian framework. First, a link is established between Bayesian estimation and inference under parameter rounding (quantization), thereby connecting two distinct paradigms: Bayesian inference and approximate computing. Next, Bayesian estimation of the means from two independent multivariate normal samples is employed to justify shrinkage estimators, i.e., means shrunk toward the pooled mean. Finally, regularized linear discriminant analysis (LDA) is considered. Various shrinkage strategies for the mean are justified from a Bayesian perspective, and novel algorithms for their computation are proposed. The proposed methods are illustrated by numerical experiments on real and simulated data.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 2 canonical work pages

  1. [1]

    Outlier detection in high dimension using regularization

    Gschwandtner, M., Filzmoser, P. Outlier detection in high dimension using regularization. In Kruse, R., et al., editors,Synergies of Soft Computing and Statistics, pages 237–244, Berlin, 2013. Springer

    2013

  2. [2]

    Guo, Y., Hastie, T., Tibshirani, R. (2007). Regularized linear discriminant analysis and its application in microarrays.Biostatistics, 8, 86–100

    2007

  3. [3]

    Bernal, V., Bischoff, R., Guryev, V., Grzegorczyk, M., Horvatovich, P. (2019). Exact hypothesis testing for shrinkage-based Gaussian graphical mod- els.Bioinformatics, 35, 5011–5017

    2019

  4. [4]

    (2020).Large covariance and autocovariance ma- trices

    Bose, A., Bhattacharjee, M. (2020).Large covariance and autocovariance ma- trices. Boca Raton: CRC Press

    2020

  5. [5]

    Boudt, K., Rousseeuw, P.J., Vanduffel, S., Verdonck, T. (2020). The minimum regularized covariance determinant estimator.Statistics and Computing, 30, 113–128. doi:10.1007/s11222-019-09869-x

    work page doi:10.1007/s11222-019-09869-x 2020

  6. [6]

    Calvetti, D., Somersalo, E. (2018). Inverse problems: From regularization to Bayesian inference.WIREs Computational Statistics, 10, e1427

    2018

  7. [7]

    (2014).Data analysis and approximate models: Model choice, location-scale, analysis of variance, nonparametric regression and image anal- ysis

    Davies, P. (2014).Data analysis and approximate models: Model choice, location-scale, analysis of variance, nonparametric regression and image anal- ysis. Boca Raton: Chapman & Hall/CRC. doi:10.1201/b17146

    work page doi:10.1201/b17146 2014

  8. [8]

    Efron, B. (2024). Machine learning and the James–Stein estimator.Japanese Journal of Statistics and Data Science, 7, 257–266

    2024

Show all 46 references
  1. [9]

    Evans, I.G. (1965). Bayesian estimation of parameters of a multivariate normal distribution.Journal of the Royal Statistical Society B, 27, 279–283. 18Jan Kalina

    1965

  2. [10]

    Fu, R., Han, M., Tian, Y., Shi, P. (2020). Improvement motor imagery EEG classification based on sparse common spatial pattern and regularized discrim- inant analysis.Journal of Neuroscience Methods, 343, 108833

    2020

  3. [11]

    Guo, Y., Hastie, T., Tibshirani, R. (2023). rda: Shrunken centroids reg- ularized discriminant analysis. R package version 1.2-1,https://CRAN.R- project.org/package=rda

    2023

  4. [12]

    Hans, C. (2009). Bayesian lasso regression.Biometrika, 96, 835–845

    2009

  5. [13]

    Hastie, T. (2020). Ridge regularization: An essential concept in data science. Technometrics, 62, 426–433

    2020

  6. [14]

    (2015).Statistical learning with sparsity

    Hastie, T., Tibshirani, R., Wainwright, M. (2015).Statistical learning with sparsity. The lasso and generalizations. Boca Raton: CRC Press

    2015

  7. [15]

    Hoffman, J., Roberts, D.A., Yaida, S. (2019). Robust learning with Jacobian regularization. ArXiv:1908.02729

    2019 arXiv

  8. [16]

    Hosseini, S.R., Deiri, E., Jamkhaneh, E.B. (2024). The Bayesian and Bayesian shrinkage estimators under square error and Al-Bayyati loss functions with right censoring scheme.Communications in Statistics—Simulation and Com- putation, 53, 3234–3252

    2024

  9. [17]

    Hubert, M., Debruyne, M., Rousseeuw, P.J. (2018). Minimum covariance de- terminant and extensions.WIREs Computational Statistics, 10, e1421

    2018

  10. [18]

    Hubert, M., Raymaekers, J., Rousseeuw, P.J. (2024). Robust discriminant analysis.WIREs Computational Statistics, 16, e70003

    2024

  11. [19]

    James, W., Stein, C. (1961). Estimation with quadratic loss. InProceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, Vol. 1, pp. 361–379. Berkeley: University of California Press

    1961

  12. [20]

    (2022).Bayes rules! An introduction to applied Bayesian modeling

    Johnson, A.A., Ott, M.Q., Dogucu, M. (2022).Bayes rules! An introduction to applied Bayesian modeling. Boca Raton: CRC Press

    2022

  13. [21]

    Jurczyk, T. (2012). Outlier detection under multicollinearity.Journal of Sta- tistical Computation and Simulation, 82, 261–278

    2012

  14. [22]

    (2019).Robust statistical methods with R, 2nd ed

    Jureckova, J., Picek, J., Schindler, M. (2019).Robust statistical methods with R, 2nd ed. Boca Raton: CRC Press

    2019

  15. [23]

    Kalina, J. (2018). A robust pre-processing of BeadChip microarray images. Biocybernetics and Biomedical Engineering, 38, 556–563

    2018

  16. [24]

    Kalina, J., Duintjer Tebbens, J. (2015). Algorithms for regularized linear discriminant analysis. InProceedings of the 6th International Conference on Bioinformatics Models, Methods and Algorithms (BIOINFORMATICS 15), pp. 128–133. Lisbon: Scitepress. Bayesian Framework for Reg...

    2015

  17. [25]

    Kalina, J., Matonoha, C. (2020). A sparse pair-preserving centroid-based su- pervised learning method for high-dimensional biomedical data or images.Bio- cybernetics and Biomedical Engineering, 40, 774–786

    2020

  18. [26]

    Ledoit, O., Wolf, M. (2003). Honey, I shrunk the sample covariance matrix. UPF Economics and Business Working Paper 691

    2003

  19. [27]

    Ledoit, O., Wolf, M. (2022). The power of (non-)linear shrinking: A review and guide to covariance matrix estimation.Journal of Financial Econometrics, 20, 187–218

    2022

  20. [28]

    Li, C.N., Ren, P.W., Guo, Y.R., Ye, Y.F., Shao, Y.H. (2024). Regularized linear discriminant analysis based on generalized cappedℓ 2,q-norm.Annals of Operations Research, 339, 1433–1459

    2024

  21. [29]

    (2023).Stein estimation

    Maruyama, Y., Kubokawa, T., Strawderman, W.E. (2023).Stein estimation. Singapore: Springer

    2023

  22. [30]

    Nahshan, Y., Chmiel, B., Baskin, C., Zheltonozhskij, E., Banner, R., Bron- stein, A.M., Mendelson, A. (2021). Loss aware post-training quantization.Ma- chine Learning, 110, 3245–3262

    2021

  23. [31]

    Piotrowski, A.P., Napiorkowski, J.J., Piotrowska, A.E. (2020). Impact of deep learning-based dropout on shallow neural networks applied to stream temper- ature modelling.Earth-Science Reviews, 201, 103076

    2020

  24. [32]

    Ribeiro, R.R.M., Natal, J., de Campos, C.P., Maciel, C.D. (2024). Conditional probability table limit-based quantization for Bayesian networks: Model qual- ity, data fidelity and structure score.Applied Intelligence, 54, 4668–4688

    2024

  25. [33]

    Rizzelli, S., Rousseau, J., Petrone, S. (2024). Empirical Bayes in Bayesian learning: Understanding a common practice. ArXiv:2402.19036v1

    2024 arXiv

  26. [34]

    Roelant, E., Van Aelst, S., Willems, G. (2009). The minimum weighted covari- ance determinant estimator.Metrika, 70, 177–204

    2009

  27. [35]

    Saha, R., Srivastava, V., Pilanci, M. (2023). Matrix compression via random- ized low rank and low precision factorization. ArXiv:2310.11028

    2023 arXiv

  28. [36]

    (2019).Theory of ridge regression estimation with applications

    Saleh, A.K.M.E., Arashi, M., Kibria, B.M.G. (2019).Theory of ridge regression estimation with applications. Hoboken: Wiley

    2019

  29. [37]

    Schafer, J., Strimmer, K. (2005). A shrinkage approach to large-scale covari- ance matrix estimation and implications for functional genomics.Statistical Applications in Genetics and Molecular Biology, 4, 32

    2005

  30. [38]

    Sifaou, H., Kammoun, A., Alouini, M.S. (2020). High-dimensional linear dis- criminant analysis classifier for spiked covariance model.Journal of Machine Learning Research, 21, 1–24. 20Jan Kalina

    2020

  31. [39]

    Stein, C.M. (1981). Estimation of the mean of a multivariate normal distribu- tion.Annals of Statistics, 9, 1135–1151

    1981

  32. [40]

    (2020).Efficient processing of deep neural networks

    Sze, V., Chen, Y.H., Yang, T.J., Emer, J.S. (2020).Efficient processing of deep neural networks. San Rafael: Morgan & Claypool

    2020

  33. [41]

    Tibshirani, R., Hastie, T., Narasimhan, B. (2003). Class prediction by nearest shrunken centroids, with applications to DNA microarrays.Statistical Science, 18, 104–117

    2003

  34. [42]

    Valenta, Z., Kalina, J. (2015). Exploiting Stein’s paradox in analysing sparse data from genome-wide association studies.Biocybernetics and Biomedical En- gineering, 35, 64–67

    2015

  35. [43]

    (2023).Computational linear algebra with applications and MAT- LAB computations

    White, R.E. (2023).Computational linear algebra with applications and MAT- LAB computations. Boca Raton: CRC Press

    2023

  36. [44]

    (2013).Robust data mining

    Xanthopoulos, P., Pardalos, P.M., Trafalis, T.B. (2013).Robust data mining. New York: Springer

    2013

  37. [45]

    Xie, P., Xiang, H., Wei, Y. (2019). Randomized algorithms for total least squares problems.Numerical Linear Algebra with Applications, 26, e2219

    2019

  38. [46]

    Yan, X., Song, J., An, Y., Lu, W. (2024). Uncertainty quantization of me- teorological input and model parameters for hydrological modelling using a Bayesian-based integrated approach.Hydrological Processes, 38, e15040

    2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.