Pith. sign in

REVIEW 5 major objections 6 minor 26 references

GLASD: A Loss-Function-Agnostic Global Optimizer for Robust Correlation Estimation under Data Contamination and Heavy Tails

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that GLASD, a gradient-free optimizer with a logarithmic cooling schedule for uphill moves, converges almost surely to the global minimum of any continuous loss on the correlation-matrix manifold.

desk verdict Solid empirical package, but the advertised global-convergence theorem is unproven—still worth referee time for the method. read the letter →

arxiv 2506.14801 v1 pith:IHPNYCWT submitted 2025-06-02 stat.AP math.OCstat.CO

Priyam Das This is my paper · ORCID
classification stat.APmath.OCstat.CO MSC 62F3562H2090C2665K05
keywords globaloptimizationrobustcorrelationestimationnon-convexblack-boxsimulatedannealingCholeskyparameterizationcontaminateddataheavytails
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

The paper introduces GLASD, a gradient-free, black-box optimizer that searches over the space of correlation matrices by way of a bijective map to a hyperrectangle. It claims that GLASD converges almost surely to the global minimum of any continuous loss on that domain, without convexity or smoothness assumptions. Because robust correlation losses such as truncated Mahalanobis and Tukey's biweight are non-convex and non-differentiable, the claim would replace bespoke robust estimators with one general-purpose solver. The paper supports the claim with a convergence theorem, benchmark studies on non-convex test functions, and simulations with contaminated and heavy-tailed data. A careful reader will note that the proof invokes a logarithmic cooling-schedule theorem whose acceptance-probability conditions differ from GLASD's actual acceptance rule.

What carries the argument

The load-bearing mechanism is the forced exploration step with logarithmic acceptance probability: with probability $1/m$ per iteration, GLASD samples a coordinate, a sign, and a magnitude from $\mathrm{U}(0,r)$, accepts improving moves always, and accepts non-improving moves with probability $q_t = \min(1, mc/\log(1+t))$. This is what the convergence theorem relies on. The second mechanism is the bijective parameterization of full-rank correlation matrices: via the Cholesky factor $L$ with unit-norm rows, the matrix is encoded by angular coordinates (the row vectors lie on unit hemispheres), mapping the correlation-manifold constraint to an open hyperrectangle in Euclidean space. That map makes GLASD's coordinate-wise box-constrained search directly applicable to correlation matrices. The algorithm's adaptive step sizes and direction probabilities ($s_{\text{inc}}, s_{\text{dec}}, p_{\text{inc}}, p_{\text{dec}}$) are secondary tuning that the theorem does not use.

What would settle it

Run GLASD on a one-dimensional step function with a narrow well containing the global minimum; if it fails to find that well with probability approaching one over many restarts and long runs, then the theorem's continuity assumption or its cooling-schedule hypothesis does not hold.

Watch

Extended reading notes

Core claim

The central claim is that GLASD, whose forced-exploration uphill moves are accepted with probability $\min(1, mc/\log(1+t))$, converges almost surely to the set of global minimizers of any continuous objective defined on a compact hyperrectangle, and hence on the correlation-matrix manifold via the paper's Cholesky-hyperspherical parameterization. The algorithm alternates between coordinate-wise greedy descent, which adapts step sizes and direction probabilities, and a forced exploration mode that samples a coordinate, sign, and step magnitude uniformly. The theorem asserts that because the exploration proposal kernel covers every axis-aligned neighborhood and uphill moves occur with probability at least $c/\log(1+t)$, the induced inhomogeneous Markov chain converges in probability to the global minimizer set. The paper also proves that the deterministic ASD variant, without exploration, converges linearly under strong convexity and almost surely under general convexity. On the applied side, the paper claims Huber, truncated, and Tukey losses, optimized by GLASD, produce lower RMSE than Gaussian likelihood under row, column, and random contamination and heavy tails.

Load-bearing premise

The convergence proof assumes that a logarithmic cooling-schedule theorem designed for temperature-based acceptance applies to GLASD's acceptance rule, which ignores how much worse the new point is, and that the loss is continuous even though the motivating truncated and Tukey losses are discontinuous.

Editorial extensions

If this is right

  • If Theorem 1 holds, a single black-box optimizer can minimize any continuous user-defined loss over correlation matrices, removing the need for bespoke solvers per loss function.
  • The Cholesky-hyperspherical parameterization provides a projection-free way to enforce symmetry, positive definiteness, and unit diagonals during optimization, which carries over to other constrained matrix estimation tasks.
  • The benchmark and simulation results, if representative, imply that GLASD with Huber or truncated losses outperforms Gaussian-likelihood estimation under row, column, and random contamination and heavy-tailed $t$-distributions.
  • The almost-sure convergence of the exploration-free ASD variant under convexity suggests that when the loss is convex, GLASD's exploration is unnecessary for convergence, though its rate guarantee is lost.

Reading between the lines

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

  • A testable extension is whether replacing the energy-independent acceptance probability $q_t$ with an energy-dependent Metropolis acceptance, such as $\min(1, \exp(-\Delta/T_t))$ with $T_t$ proportional to $1/\log t$, would satisfy the standard logarithmic cooling hypotheses and extend the proof to discontinuous losses.
  • If the continuity assumption in Theorem 1 is the actual barrier, the motivating truncated and Tukey losses are discontinuous, so practitioners should verify that the reported empirical convergence persists for discontinuous objectives; a counterexample on a step function would delimit the theorem's reach.
  • The parameterization may transfer to other structured matrices, such as precision matrices or low-rank factors, where unit-norm row constraints appear naturally; testing GLASD on those manifolds would tell whether the framework is genuinely loss-agnostic or specific to correlation matrices.
  • The observed superiority of Huber over Tukey in the simulations suggests that loss choice still matters even with a global optimizer; a fair comparison would report the sensitivity of RMSE to the IQR-based threshold settings for each loss.
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

5 major / 6 minor

Summary. The manuscript introduces GLASD (Global Adaptive Stochastic Descent), a gradient-free, black-box optimizer for loss functions defined over the space of correlation matrices, with the goal of unifying robust correlation estimation under arbitrary user-specified losses. The paper claims almost-sure convergence of GLASD to the global minimizer set for continuous objectives on compact hyperrectangles (Theorem 1), linear and almost-sure convergence results for its exploration-free variant ASD under convexity (Theorems 2 and 3), and a bijective hyperspherical parameterization of full-rank correlation matrices (Theorem 4). The method is evaluated on benchmark functions over the correlation manifold, in simulation studies under four contamination scenarios and three correlation structures, and on a breast cancer proteomic data set. The central theoretical claim is that GLASD is a loss-function-agnostic global optimizer, with supporting open-source MATLAB code.

Significance. If the global-convergence result were valid, the paper would offer a practically appealing contribution: a single black-box optimizer that handles nonconvex, nonsmooth, and even discontinuous losses over the correlation-matrix manifold without gradients, backed by reproducible open-source software. The simulation design is extensive, covering multiple contamination types, correlation structures, and dimensions, and the real-data example illustrates a plausible use case. However, the theoretical core is not sound: the proof of Theorem 1 invokes a simulated-annealing convergence theorem whose hypotheses are not met by GLASD's acceptance rule, the compactness assumption fails for the intended correlation-matrix application, and Theorems 2 and 3 contain load-bearing mathematical errors. The empirical benchmark results also do not support the global-convergence claim. The paper's strengths are its reproducible code, thorough empirical coverage, and clear exposition; its central claimed guarantee is unsupported.

major comments (5)
  1. [Section IV, Theorem 1 and its proof] The proof invokes Hajek's logarithmic cooling theorem [14] to conclude almost-sure convergence to the global minimizer set, but GLASD's uphill acceptance probability q_t = min(1, mc/log(1+t)) is independent of the objective increase f(y)-f(x). Hajek's theorem applies to time-inhomogeneous Markov chains whose transition kernels admit a Gibbs stationary law proportional to exp(-f/T_t), which requires acceptance probabilities of the form exp(-(f(y)-f(x))/T_t). Since GLASD's acceptance rule is not of this form, the chain need not have a Gibbs invariant law, and the claimed almost-sure convergence does not follow. Additionally, the theorem states almost-sure convergence, but the proof concludes only lim_{t→∞} P(x_t ∈ X*) = 1, which is convergence in probability, not almost-sure convergence.
  2. [Section V, Theorem 4 and Section IV, Theorem 1] The compactness hypothesis of Theorem 1 is not satisfied in the correlation-matrix application. Theorem 4 maps full-rank correlation matrices to an open hyperrectangle: for example, ω21 ∈ (-π/2, π/2) and ω_m1 ∈ (0, π/2), so l22 > 0 and l_mm > 0. The GLASD algorithm is defined on a compact hyperrectangle D = ∏[a_i,b_i], and the motivating losses in Section II, such as ℓGauss in Eq. (1), contain log det(C), which diverges to -∞ as C approaches the boundary of the correlation-matrix space. Thus f need not be continuous on the closed domain, and the assumptions of Theorem 1 are not met for the intended application.
  3. [Section IV, Theorem 2 proof] The inequality f(x + s_j v_j) ≤ f(x) - (s_j^2/(2L))(∇f(x)^T v_j)^2 does not follow from L-smoothness for an arbitrary step size s_j; it holds only when s_j equals the optimal step α* = -∇f(x)^T v_j / L. For example, with f(x) = x^2/2, L = 1, x = 1, v_j = 1, and s_j = 0.1, the left-hand side is f(1.1) = 0.605, while the right-hand side is 0.5 - 0.005 = 0.495. Hence the linear convergence rate claimed in the theorem is unproven. Also, condition (d), 'Exact gradients are available,' conflicts with the paper's characterization of ASD as a gradient-free algorithm.
  4. [Section IV, Theorem 3 proof] The proof argues that every limit point x∞ of the ASD sequence satisfies f(x∞) = inf_t f(x_t), and then claims that if the minimizer set is nonempty and x∞ is not a minimizer, this 'contradicts the assumption that x∞ is not a minimizer' because 'f(x∞) = inf_t f(x_t) → f*'. The expression inf_t f(x_t) is a constant, not a sequence, and the fact that the monotone sequence f(x_t) converges to f∞ does not imply f∞ = f*. The conclusion that all limit points lie in the minimizer set and that dist(x_t, X*) → 0 almost surely is therefore not established by the proof as written.
  5. [Section VI, Table I] The reported benchmark results contradict the global-convergence claim. For the Rastrigin benchmark, the paper states that the global minimum is attained at the identity matrix, where the objective value is 0, yet GLASD's best values are 1.01E+01, 1.13E+02, 6.83E+02, and 3.40E+03 for M = 5, 10, 20, 50. Several reported standard errors are also very large (e.g., Rosenbrock M=5 standard error 1.40E+08), indicating high run-to-run variability. Even with finite computational budgets, these results do not support the statement that GLASD consistently behaves as a global optimizer.
minor comments (6)
  1. [Abstract and Section I] The abstract and introduction claim global convergence guarantees under mild regularity conditions and the ability to handle discontinuous objectives, but Theorem 1 requires continuity and compactness; this discrepancy should be clarified.
  2. [Section IV, Theorem 1 proof] The proof states that the proposal kernel has support covering every axis-aligned neighborhood, but a single coordinate step changes only one coordinate; the claimed ψ-irreducibility with respect to Lebesgue measure needs a more careful multi-step argument.
  3. [Section V] The notation M is used both for the correlation-matrix dimension and for the stagnation window in Algorithm 1; the parenthetical clarification in Section V is confusing and should be resolved with distinct symbols.
  4. [Section V, tuning parameters] The default cooling constant c = 0.001 log(n) depends on the problem dimension n, while Theorem 1 treats c as a fixed constant for a given problem; the relationship between the default and the theoretical condition should be discussed.
  5. [Algorithm 1] The step-size and probability updates are executed only when an exploration step is not involved; the asymmetry between accepted greedy steps and accepted exploration steps is not discussed and may affect the adaptive behavior of the algorithm.
  6. [Section VIII] The case study reports no quantitative comparison with a non-robust baseline, such as the sample correlation matrix; the text acknowledges this, but adding such a comparison would make the robustness benefit more concrete.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity: the central convergence proof fails because Hajek's theorem does not apply to GLASD's energy-independent acceptance probability, but that is a soundness error, not a self-referential one.

full rationale

GLASD's central claim is a convergence theorem, not a fitted prediction. The only self-citation ([26]) selects the breast-cancer protein subset in the case study and plays no role in the proof of Theorem 1 or in the simulation benchmarks; it is therefore not load-bearing. The acceptance probability q_t = min(1, mc/log(1+t)) is an algorithmic design choice, not a parameter fitted to the global-minimizer set, so no fitted-input-called-prediction pattern arises. The parameterization in Theorem 4 is standard spherical-coordinate decomposition of the Cholesky factor and is referred to an external source for the bijection [24]; it is not used to smuggle in the convergence conclusion. The serious flaw in Section IV is that the proof invokes Hajek's logarithmic-cooling theorem [14] although that theorem requires acceptance probabilities of Gibbs form exp(-Delta/T) satisfying detailed balance, whereas GLASD's uphill acceptance probability does not depend on f(y)-f(x); consequently the global-convergence guarantee is unsupported. This is a soundness error, not a circularity: the theorem statement is not defined in terms of its own conclusion, no equation reduces to the target result by construction, and the failure is a mismatch between the hypotheses of the cited theorem and the algorithm, not a self-referential chain. The paper does not derive its central result from the author's own prior work. Accordingly, no circular step is identified; the score reflects only the non-load-bearing self-citation, not circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central convergence theorems rest on several unverified assumptions: continuity despite discontinuous losses, applicability of Hajek's theorem to a time-independent acceptance rule, and a false smoothness inequality in Theorem 2. The standard hyperspherical parameterization is known mathematics. No new physical or modeling entities are introduced.

free parameters (6)
  • initial step sizes s_j = 0.1
    Default step size chosen by hand in Section V; GLASD's search behavior depends on it.
  • step and probability multipliers sinc, sdec, pinc, pdec = 2 each
    Hand-chosen adaptation rates in Algorithm 1; no sensitivity analysis is given.
  • exploration frequency m = 5
    Hand-chosen in Section V; controls the balance between greedy and forced exploration.
  • cooling constant c = 0.001 log(n)
    Hand-chosen; controls the uphill acceptance probability and is essential to the claimed global convergence.
  • exploration radius r = distance to domain boundary
    Set dynamically in Section V; affects the size of exploration steps.
  • robust loss thresholds delta and tau = Q3 + 3 IQR
    Chosen by an IQR-based rule in Section VII; the RMSE results for Huber, truncated, and Tukey losses depend on this choice.
assumptions (5)
  • domain assumption The objective f is continuous and the domain D is compact (Theorem 1).
    Used in the proof of Theorem 1, but Section II motivates discontinuous truncated and Tukey losses, and the spherical parameterization is an open hyperrectangle, so the assumption does not cover the headline use case.
  • ad hoc to paper Hajek's logarithmic cooling theorem applies to a Markov chain whose uphill acceptance probability is independent of the energy difference.
    Invoked in Section IV, proof of Theorem 1, without verification. Standard annealing results require acceptance probabilities tied to objective-value differences, so this premise is unsupported.
  • domain assumption A coordinate-wise proposal distribution with clipping is psi-irreducible with respect to Lebesgue measure on the whole domain.
    Used in Theorem 1's proof; the time-dependent acceptance rule may not yield the recurrence conditions required by Hajek's theorem.
  • ad hoc to paper L-smoothness and strong convexity, together with strict-decrease coordinate steps, yield a decrease lower bound proportional to the squared gradient norm.
    This is the false inequality in the proof of Theorem 2; without it the claimed linear rate does not follow.
  • standard math The hyperspherical Cholesky parameterization is bijective onto the full set of full-rank correlation matrices.
    Theorem 4 states a standard result, but the paper does not address the mismatch between the open parameter domain and the compact domain required by Theorem 1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GLASD: A Loss-Function-Agnostic Global Optimizer for Robust Correlation Estimation under Data Contamination and Heavy Tails." pith.science (2026). https://pith.science/paper/IHPNYCWT

@misc{pith2026250614801,
  author       = {Pith},
  title        = {Pith review of: GLASD: A Loss-Function-Agnostic Global Optimizer for Robust Correlation Estimation under Data Contamination and Heavy Tails},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IHPNYCWT}},
  note         = {Machine review of arXiv:2506.14801}
}
read the original abstract

Robust correlation estimation is essential in high-dimensional settings, particularly when data are contaminated by outliers or exhibit heavy-tailed behavior. Many robust loss functions of practical interest-such as those involving truncation or redescending M-estimators-lead to objective functions that are inherently non-convex and non-differentiable. Traditional methods typically focus on a single loss function tailored to a specific contamination model and develop custom algorithms tightly coupled with that loss, limiting generality and adaptability. We introduce GLASD (Global Adaptive Stochastic Descent), a general-purpose black-box optimization algorithm designed to operate over the manifold of positive definite correlation matrices. Unlike conventional solvers, GLASD requires no gradient information and imposes no assumptions of convexity or smoothness, making it ideally suited for optimizing a wide class of loss functions-including non-convex, non-differentiable, or discontinuous objectives. This flexibility allows GLASD to serve as a unified framework for robust estimation under arbitrary user-defined criteria. We demonstrate its effectiveness through extensive simulations involving contaminated and heavy-tailed distributions, as well as a real-data application to breast cancer proteomic network inference, where GLASD successfully identifies biologically plausible interactions despite the presence of outliers. The proposed method is scalable, constraint-aware, and available as open-source software at GitHub.

Figures

Figures reproduced from arXiv: 2506.14801 by the authors.

Figure 1
Figure 1. Convergence profiles of GLASD and competing optimization methods on benchmark functions. Log-scaled objective values are plotted against the [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure: Protein-level expression characteristics in four major activated [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 23 canonical work pages

  1. [14]

    Cooling schedules for optimal annealing,

    B. Hajek, “Cooling schedules for optimal annealing,” Mathematics of Operations Research, vol. 13, no. 2, pp. 311–329, 1988

    work page 1988

  2. [1]

    A framework for robust high-dimensional correlation estimation in genomics,

    A. T. Shafer and R. P. Adams, “A framework for robust high-dimensional correlation estimation in genomics,” Bioinformatics, vol. 31, no. 9, pp. 1452–1460, 2015

    work page 2015

  3. [2]

    Covariance regularization by thresholding,

    P. J. Bickel and E. Levina, “Covariance regularization by thresholding,” Ann. Statist., vol. 36, no. 6, pp. 2577–2604, 2008

    2008

  4. [3]

    Network modelling methods for FMRI,

    S. M. Smith et al., “Network modelling methods for FMRI,” NeuroImage, vol. 54, no. 2, pp. 875–891, 2011

    work page 2011

  5. [4]

    An overview of the estimation of large covariance and precision matrices,

    J. Fan, Y . Liao, and H. Liu, “An overview of the estimation of large covariance and precision matrices,” The Econometric Journal , vol. 19, no. 1, pp. C1–C32, 2016

    work page 2016

  6. [5]

    Robust mean and eigenvalue regularized covariance matrix estimation,

    C. Lam, “Robust mean and eigenvalue regularized covariance matrix estimation,” Electronic Journal of Statistics , vol. 10, no. 2, pp. 4059– 4094, 2016

    work page 2016

  7. [6]

    Robust estimation of a location parameter,

    P. J. Huber, “Robust estimation of a location parameter,” Ann. Math. Statist., vol. 35, no. 1, pp. 73–101, 1964

    1964

  8. [7]

    Robust covariance estimation for approximate factor models,

    J. Fan, W. Wang, and Y . Zhong, “Robust covariance estimation for approximate factor models,” Journal of Econometrics , vol. 208, no. 1, pp. 5–22, 2019

    work page 2019

Show all 26 references
  1. [8]

    The fitting of power series, meaning polynomi- als, illustrated on band-spectroscopic data,

    A. Beaton and J. Tukey, “The fitting of power series, meaning polynomi- als, illustrated on band-spectroscopic data,” Technometrics, vol. 16, no. 2, pp. 147–185, 1974

    1974

  2. [9]

    A robust measure of correlation between two genes on a microarray,

    J. Hardin, A. Mitani, L. Hicks, and M. VanKoten, “A robust measure of correlation between two genes on a microarray,” BMC Bioinformatics , vol. 8, no. 220, 2007

    2007

  3. [10]

    Absil, R

    P.-A. Absil, R. Mahony, and R. Sepulchre, Optimization Algorithms on Matrix Manifolds. Princeton, NJ, USA: Princeton University Press, 2009

    2009

  4. [11]

    Boumal, An Introduction to Optimization on Smooth Manifolds

    N. Boumal, An Introduction to Optimization on Smooth Manifolds . Cambridge, U.K.: Cambridge University Press, 2023

    2023

  5. [12]

    Optimization by adaptive stochastic descent,

    C. C. Kerr, S. Dura-Bernal, T. G. Smolinski, G. L. Chadderdon, and D. P. Wilson, “Optimization by adaptive stochastic descent,” PLOS ONE, vol. 13, no. 3, pp. e0192944, 2018

    2018

  6. [13]

    Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images,

    S. Geman and D. Geman, “Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. PAMI-6, no. 6, pp. 721–741, 1984

    1984

  7. [15]

    TCPA: a resource for cancer functional proteomics data,

    J. Li et al., “TCPA: a resource for cancer functional proteomics data,” Nat. Methods, vol. 10, no. 11, pp. 1046–1047, 2013

    2013

  8. [16]

    Virtual Library of Simulation Ex- periments: Test Functions and Datasets,

    S. Surjanovic and D. Bingham, “Virtual Library of Simulation Ex- periments: Test Functions and Datasets,” 2013. [Online]. Available: https://www.sfu.ca/∼ssurjano/optimization.html

    2013

  9. [17]

    Test functions for optimization needs,

    M. Molga and C. Smutnicki, “Test functions for optimization needs,” Tech. Rep., Poznan University of Technology, 2005

    2005

  10. [18]

    Optimization by simulated annealing,

    S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by simulated annealing,” Science, vol. 220, no. 4598, pp. 671–680, 1983

    1983

  11. [19]

    J. H. Holland, Adaptation in Natural and Artificial Systems . Cambridge, MA: MIT Press, 1992

    1992

  12. [20]

    Particle swarm optimization,

    J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proc. IEEE Int. Conf. Neural Networks , 1995, vol. 4, pp. 1942–1948

    1995

  13. [21]

    A trust region method based on interior point techniques for nonlinear programming,

    R. H. Byrd, J. C. Gilbert, and J. Nocedal, “A trust region method based on interior point techniques for nonlinear programming,” Mathematical Programming, vol. 89, no. 1, pp. 149–185, 2000

    2000

  14. [22]

    Nocedal and S

    J. Nocedal and S. Wright, Numerical Optimization, 2nd ed. New York, NY , USA: Springer, 2006

    2006

  15. [23]

    On the convergence of pattern search algorithms,

    V . Torczon, “On the convergence of pattern search algorithms,” SIAM Journal on Optimization , vol. 7, no. 1, pp. 1–25, 1997

    1997

  16. [24]

    A derivation of n-dimensional spherical coordinates,

    L. E. Blumenson, “A derivation of n-dimensional spherical coordinates,” Amer. Math. Monthly, vol. 67, no. 1, pp. 63–66, 1960

    1960

  17. [25]

    Manopt, a Matlab Toolbox for Optimization on Manifolds

    N. Boumal, B. Mishra, P.A. Absil, and R. Sepulchre (2014). “Manopt, a Matlab Toolbox for Optimization on Manifolds.”, Journal of Machine Learning Research, vol. 15, no. 42, pp. 1455–1459, 2014

    2014

  18. [26]

    NExUS: Bayesian simultaneous network estimation across unequal sample sizes,

    P. Das, C. B. Peterson, K. A. Do, R. Akbani, and V . Baladandayuthapani, “NExUS: Bayesian simultaneous network estimation across unequal sample sizes,” Bioinformatics, vol. 36, no. 3, pp. 798–804, 2020. Priyam Das received the B.Stat. and M.Stat. degrees in statistics from the...

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.