Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Simulation-based Inference via Langevin Dynamics with Score Matching

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Training a debiased, curvature-regularized score network to drive Langevin Monte Carlo makes simulation-based inference scale to moderate-dimensional parameters with one to two orders of magnitude fewer simulations.

desk verdict Solid, honest SBI methods paper with a real new combination—the central theorem rests on an assumed uniform error rate that is neither proved nor measured, but it deserves serious refereeing. read the letter →

arxiv 2509.03853 v3 pith:J2O3ZJCX submitted 2025-09-04 stat.ME stat.COstat.ML

Haoyu Jiang , Yuexi Wang , Yun Yang This is my paper · ORCID
classification stat.MEstat.COstat.ML
keywords BayesianInferenceMonteCarloMethodsLangevinDynamicsSamplingAlgorithmScoreMatchingSimulation-based
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 tries to show that simulation-based Bayesian inference need not stay stuck in low dimensions or demand tens of millions of simulator calls. It trains a score network to approximate the gradient of the log-likelihood from simulated data, then uses that estimated score inside Langevin Monte Carlo to sample the posterior. The central claim is that a localized, debiased, curvature-regularized single-data score estimator keeps the posterior approximation error growing only linearly with sample size, and that this yields tighter credible intervals than ABC, BSL, and NPE at far lower simulation cost. If true, SBI becomes practical for models with roughly a dozen parameters and large datasets where likelihoods are intractable.

What carries the argument

The central object is the debiased, structurally regularized score network: a neural network s_phi(theta, x) trained by score matching to approximate the per-observation likelihood score, with the full-data score formed by summation over observations. It embeds three properties every regular parametric likelihood satisfies: additivity across observations, the mean-zero property E[s*]=0, and the curvature identity E[s* s*^T + ∇s*]=0, enforced by a curvature-matching penalty and a debiasing step that subtracts a fitted mean. Localization via sliced Wasserstein distance builds a proposal q(theta) concentrated near the true parameter. The machinery works by making the uniform score, curvature, a

What would settle it

Run the proposed algorithm on the monotonic regression example and directly measure the three uniform error quantities in Assumption 9 over the localized set A_{n,1} for increasing n. If the single-data score error does not decay at least as 1/log n, or the mean error does not go to zero, then Theorem 3's score-error term cannot vanish and the claimed posterior approximation guarantee fails.

Watch

Extended reading notes

Core claim

The paper claims that scalable simulation-based inference can be achieved by replacing intractable likelihood evaluations with a learned score function and using it to run Langevin Monte Carlo. The key mechanism is a debiased single-data score network: the full-data score is written as a sum over i.i.d. observations, the network is constrained to satisfy the mean-zero and Fisher-curvature identities of likelihood scores, and a fitted mean is subtracted to prevent cumulative score bias from growing quadratically with n. A localization step, built on sliced Wasserstein distance matching, first concentrates the training distribution in a root-n neighborhood of the true parameter. Under regulari

Load-bearing premise

Everything rests on the trained score network being accurate enough—in its predictions, their curvature, and their mean—throughout the narrow region where the posterior actually lives, and on the localization step actually finding that region.

Editorial extensions

If this is right

  • If the central claim holds, simulation-based inference for moderate-dimensional parameters (roughly 7–12 parameters) becomes feasible with 10^4–10^5 simulation units, compared with roughly 10^6 for ABC, BSL, and NPE in the paper's experiments.
  • The debiasing step converts what would otherwise be quadratic growth of the full-data score error into linear growth, so the single-data score-matching approach can work with only a 1/log n single-observation score error rate.
  • Localization reduces the effective training domain from O(1) to O(n^{-1/2}) and the additive structure cuts the score network's input dimension from d_theta + np to d_theta + p, improving scalability and approximation rates.
  • Because the learned score is gradient information, the same structural regularization could be plugged into other gradient-based samplers such as Hamiltonian Monte Carlo or preconditioned Langevin dynamics, as the paper notes in its discussion.
  • The theoretical bound explicitly separates burn-in, discretization, and score errors, giving practitioners concrete guidance on step size (order 1/n) and concentration of the initial distribution.

Reading between the lines

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

  • The paper's theory assumes the trained network achieves the three uniform error rates in Assumption 9, but it never measures those rates on its own networks; a natural check is to estimate these error terms directly on the localized set and see whether the observed posterior errors track the bound.
  • The debiasing lemma—subtracting a fitted mean never increases score-matching error—is generic and could likely improve other amortized score or likelihood estimators, not only this Langevin sampler, though the paper does not test that.
  • The root-n localization guarantee relies specifically on sliced Wasserstein distance and Assumptions 3–4; replacing it with a cheaper or non-metric discrepancy would require a new concentration argument, and one could test whether MMD- or Euclidean-based localization preserves the reported interval tightness on the epidemic model.
  • The reported benefits are demonstrated for unimodal posteriors; extending to multimodal posteriors would require mode discovery or annealing, and the paper does not quantify how the structural constraints behave across multiple modes.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes an SBI method that combines a localized proposal distribution, score matching with structural regularizations (additivity, curvature, mean-zero debiasing), and Langevin Monte Carlo sampling. The main theoretical results are a convergence rate for the localization step (Theorem 2) and posterior approximation error bounds for both full-data score matching (Theorem 5) and the proposed debiased single-data score matching (Theorem 3). The key algorithmic idea is to estimate a single-observation score network, debias it by subtracting its fitted mean, and then sum the debiased scores over the n observations, so that the cumulative score error grows linearly rather than quadratically in n. Empirically, the method is compared with ABC, BSL, and NPE on an M/G/1 queueing model, a Bayesian monotonic regression model, and a stochastic epidemic model, reporting tighter credible intervals at comparable or reduced simulation cost.

Significance. If the theoretical guarantees were fully established, this would be a valuable contribution: the additive score structure reduces simulation cost from O(N n) to O(N), the curvature regularization gives local geometric accuracy for LMC, and the debiasing step offers a clean bias-variance decomposition that keeps the full-data score error linear in n. The proofs in the appendix are systematic and, conditional on the stated assumptions, the derivations are coherent. The empirical results are promising and use external baselines. However, the central theoretical claim rests on Assumption 9 (and its full-data analogue Assumption 13), which postulates uniform sup-norm score, curvature, and mean errors over a localized parameter set A_{n,1}. This assumption is neither derived from the training losses nor verified empirically, and the cited nonparametric rates are L2 rates that do not imply the required sup-norm control. The paper's theoretical contribution is therefore currently conditional on a load-bearing premise that is not supported by the training procedure.

major comments (4)
  1. [§4.2, Assumption 9; Appendix A.5, Lemma 4] Assumption 9 defines eε^2_{N,1}, eε^2_{NR,mR,2}, and eε^2_{NR,mR,3} as suprema over A_{n,1}. The training objectives (6), (24), (7), and (25) are expectations under the proposal q(θ) (with empirical averages); no term controls the uniform sup-norm over A_{n,1}. Lemma 4 uses these suprema to bound the fourth moment of the score error, so the ε_n term in Theorem 3 is entirely a postulate. Remark 3 cites nonparametric regression L2 rates under the training measure (Shen et al. 2020; Schmidt-Hieber 2020), which do not imply sup-norm control on A_{n,1}. The same issue applies to Assumption 13 and Theorem 5. The central vanishing-error claim is therefore not established. Please either derive sup-norm rates for the proposed training procedure, reformulate the theorem around an error measure that the loss actually controls and adapt the proof, or empirically measure the three error terms on the
  2. [§3.1, Theorem 2; §4.2, A_{n,1}] The localization step builds q as N(bµ, DiagCov(bθ_b)) with B=100. Theorem 2 only guarantees each bθ_b is Op(n^{-1/2} + m^{-1/2}); hence q puts its mass within O(1) of θ* in the rescaled coordinate α=√n(θ−θ*). The set A_{n,1} is a ball of radius C0√log n, so its peripheral region has q-mass of order n^{-C0^2/2}. Score, curvature, and mean errors can be arbitrarily large on that peripheral region without affecting the q-weighted training loss. This explains why the L2-under-q rates cited in Remark 3 cannot deliver Assumption 9. The localization theory also does not quantify the finite-B effect on the empirical covariance, which affects whether q over- or under-covers A_{n,1}. This is a second, related gap in the theoretical support for Theorem 3.
  3. [§4.2, text after Theorem 3] The text states that eε_NR,mR,2 and eε_NR,mR,3 scale as O(1/√m_R) and that controlling the corresponding terms requires m_R = O(n log n). However, substituting eε_NR,mR,3 = O(1/√m_R) into ε_n^2 yields n eε^2_{NR,mR,3} log n = O(n·(1/(n log n))·log n) = O(1), not o(1). To make ε_n vanish one needs m_R = ω(n log n), e.g., m_R = n(log n)^{1+δ}. The stated sample-size prescription is therefore inconsistent with the claimed vanishing score error.
  4. [§4.2, Remark 3] Remark 3 concludes that with N=n the score-matching error is n^{-1/2} 'for both approaches'. For the non-debiased single-data analysis, the text above requires eε_{N,1} = O(1/(√n log n)). Since n^{-1/2} is larger than that bound, the term n eε^2_{N,1}(log n)^2 would be (log n)^2, not o(1). Thus the remark overstates the theoretical support for the non-debiased variant. Please clarify which error normalization is being discussed and correct the claimed sufficiency.
minor comments (6)
  1. [§2, first paragraph] Duplicate phrase: 'drawn from drawn from a distribution'.
  2. [Theorem 3 statement] The target posterior is written as π(· | X*_n) in Theorem 3, while elsewhere it is π_n(· | X*_n). Please unify the notation.
  3. [General] No code or data availability statement is provided. Given the implementation-heavy details in Appendix C, a reproducibility statement would strengthen the paper.
  4. [Figure 1] The rightmost panel has no axis labels; the caption refers to θ0 and θ1, but the axes are unlabeled.
  5. [§5.2] 'the final statistics is averaged' should be 'the final statistics are averaged'.
  6. [Appendix C.1.1, Solution 2] The sentence about the standard deviation of all dimensions of X*_n is ambiguous; it likely refers to a pooled/overall standard deviation. Please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 3 is a conditional bound in terms of explicitly assumed score-estimation errors, and the empirical comparisons are against external baselines.

full rationale

The derivation chain is not circular. The central result, Theorem 3, states a posterior approximation bound that is conditional on Assumptions 5–9, including the uniform score-, curvature-, and mean-matching errors e_eps_{N,1}, e_eps_{NR,mR,2}, e_eps_{NR,mR,3} defined over the localized set A_{n,1}. These quantities are inputs to the theorem, not outputs fitted to the reported posterior intervals or coverages. No constant in Theorem 2 or Theorem 3 is tuned to Tables 1–3; the empirical baselines (ABC, BSL, NPE) are external methods with their own fitted or tuned components, so the improvements reported are not forced by construction. The debiasing Lemma 1 is a genuine inequality: because h_psi(theta) = 0 is feasible in (7), subtracting the fitted mean cannot increase the single-data score-matching loss; this is not a renaming of the target. The localization step is a computational device that estimates a proposal q(theta) from simulations and the observed data; using this proposal to train the score network is a standard amortized-SBI construction, and Theorem 2's guarantee about q is a separate asymptotic statement, not an assumption that the final posterior equals the localization output. Cited prior work on Langevin discretization and Girsanov bounds (Ding et al. 2024; Chen et al. 2023; Tang and Yang 2024) is used as external support for standard sampling inequalities, not as a self-citation chain that determines the paper's main posterior bound. The reviewer-identified gap—that Assumption 9 requires sup-norm error over A_{n,1} while Remark 3 quotes integrated L2 nonparametric rates—is a correctness/verification concern about whether Assumption 9 is satisfied by the proposed training procedure, not a circularity: the bound explicitly depends on the assumed sup-norm errors and does not redefine those errors as the achieved L2 rate. Because circularity requires showing that a claimed prediction or derivation reduces by definition or by fitted input to the paper's own outputs, and no such reduction is present, the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 8 assumptions · 0 invented entities

Everything rests on three stacks. (1) The score-matching objective is valid only under the Hyvarinen-Dayan boundary condition (Assumption 1), which the paper admits is violated by the queueing simulator and patches with a weight function or Gaussian noise, both approximate. (2) The localization analysis needs the simulator to be reparametrizable, the inverse parameter map to be Lipschitz in sliced Wasserstein distance (Assumption 3), and uniform convergence of the empirical SW distance (Assumption 4); these are plausible but not verified for the epidemic simulator, where the indicator function must be smoothed. (3) The posterior bound needs the true posterior to concentrate and satisfy LSI (Assumptions 5, 7) and, crucially, Assumption 9/13: uniform score, curvature, and mean-matching error bounds over the localized set. That assumption is the linchpin and is not derived from the proposed training objective. Free parameters: two curvature penalties selected by validation, a noise level picked heuristically as 5-10% of data spread, and localization settings (B=100, 100 projections, m=n). No invented entities.

free parameters (5)
  • lambda1 (curvature penalty strength in score matching) = 1e-8 (queueing), 1e-3 (monotonic regression), selected by validation set
    Strength of the Fisher-information curvature penalty in losses (6) and (24); chosen by validation loss, not by theory.
  • lambda2 (curvature penalty strength in mean regression) = 1e-8 (queueing), 1e-3 (monotonic regression), selected by validation set
    Strength of the curvature penalty in the debiasing objective (7) and (25); chosen by validation.
  • noise level sigma_epsilon (boundary fix) = 0.25 in the M/G/1 queueing example
    Gaussian perturbation added to data in Solution 2 for the boundary condition; set as 5-10% of the data standard deviation, a heuristic that trades posterior bias versus numerical stability (Appendix C.1.1).
  • localization settings (B, number of SW projections, m) = B=100, 100 projections, m=n
    Number of bootstrap-like estimators and sliced Wasserstein projections in the localization step; chosen for computational cost, reported in Appendix B.1.
  • LMC sampling schedules = 3 noise-injection sets in queueing; tempering 0.1 to 1.0 in monotonic regression
    Practical sampling choices that influence the reported credible intervals; described in Appendices C.1.2 and C.2.2.
assumptions (8)
  • standard math Assumption 1: boundary condition p(theta) p_n(theta)(X) s_phi(theta, X) -> 0 at the support boundary, required for the score-matching objective (Theorem 1)
    This is the standard score-matching integration-by-parts condition from Hyvarinen and Dayan (2005). The paper notes it is violated by many simulators and patches it with a weight function or Gaussian noise (Appendix B.2).
  • domain assumption Assumption 3: the inverse map P_theta -> theta is Lipschitz in sliced Wasserstein distance
    Required for the localization rate in Theorem 2; holds for location families and strongly convex exponential families, but is not verified for the epidemic simulator.
  • domain assumption Assumption 4: uniform Op(m^{-1/2}) convergence of empirical sliced Wasserstein distance over compact Theta
    Used in Theorem 2; Lemma 2 in Appendix A.2 derives it from Lipschitz and sub-Gaussian conditions on the simulator.
  • domain assumption Assumption 5: concentration of the true posterior and MLE at rate exp(-C1 n t^2)
    Standard Bernstein-von Mises type condition; referenced to Ghosal et al. (2000) and Spokoiny (2012).
  • domain assumption Assumption 7: the scaled posterior satisfies a log-Sobolev inequality
    Needed for the LMC mixing bound; mild for large n by the Bernstein-von Mises theorem, but an assumption on the target posterior.
  • ad hoc to paper Assumption 9 / 13: uniform score, curvature, and mean-matching errors are bounded at the stated rates over the localized set A_{n,1}
    The paper does not prove these rates for its trained networks; Remark 3 cites generic nonparametric regression rates. Theorems 3 and 5 reduce the whole posterior-error claim to this assumption, so it is the linchpin.
  • domain assumption Reparametrization representation of the simulator: X = tau(theta, Z) with Z ~ P_Z known
    Needed for gradient-based localization in Section 3.1; the epidemic example requires a smooth approximation of the indicator function to make gradients available.
  • domain assumption i.i.d. structure for the additive score decomposition and the mean-zero/curvature identities
    The mean-zero and Fisher identities require the support of the data not to depend on theta (footnote 1). For dependent data (epidemic model) the structural advantages are dropped.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Simulation-based Inference via Langevin Dynamics with Score Matching." pith.science (2026). https://pith.science/paper/J2O3ZJCX

@misc{pith2026250903853,
  author       = {Pith},
  title        = {Pith review of: Simulation-based Inference via Langevin Dynamics with Score Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J2O3ZJCX}},
  note         = {Machine review of arXiv:2509.03853}
}
read the original abstract

Simulation-based inference (SBI) enables Bayesian analysis when the likelihood is intractable but model simulations are available. Recent advances in statistics and machine learning, including Approximate Bayesian Computation and deep generative models, have expanded the applicability of SBI, yet these methods often face substantial computational challenges as the sample size and parameter dimension increase. In this paper, we propose a novel scalable SBI method that integrates score matching with Langevin dynamics, while explicitly exploiting the statistical structure of log-likelihood functions. Our approach combines (i) a localization scheme that concentrates computation in regions of high posterior mass and (ii) a structured score network that embeds key properties of likelihood scores, including additivity across observations and Fisher information identities. We provide theoretical and empirical evidence demonstrating that the proposed structured score-matching approach improves statistical efficiency and computational scalability, achieving competitive or superior performance compared to existing SBI methods on both benchmark and challenging problems with large sample sizes and moderate-dimensional parameter spaces.

Figures

Figures reproduced from arXiv: 2509.03853 by the authors.

Figure 1
Figure 1. Score estimation under the monotonic regression example in Section 5.2. Shown are score directions overlaid on the heatmap of the target posterior density on (θ0, θ1). From left to right: scores estimated from the proposal q(θ), scores from the true likelihood, and scores estimated from the prior π(θ). 3.2 Score network regularization based on statistical structures A distinctive feature of score-matching networks f… view at source ↗
Figure 2
Figure 2. Posterior density plot of one experiment under the M/G/1-queuing model. [PITH_FULL_IMAGE:figures/full_fig_p027_2.png] view at source ↗
Figure 3
Figure 3. Distribution of observed data and true scores in the monotonic regression example. [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: 95% credible bands of different methods in one monotonic regression experiment [PITH_FULL_IMAGE:figures/full_fig_p030_4.png]
Figure 5
Figure 5. Figure 5: Posterior densities of different methods under the 5-floor setting [PITH_FULL_IMAGE:figures/full_fig_p031_5.png]
Figure 6
Figure 6. Figure 6: Estimation errors against prior den￾sity in the Beta-Binomial example. The esti￾mation error increases significantly when the prior density is low. As shown in [PITH_FULL_IMAGE:figures/full_fig_p053_6.png]
Figure 7
Figure 7. Figure 7: Results of Algorithm 3 on the queuing model under different noise levels As a result, we recommend using Solution 1 when the data has strong dependency structure or the weight function can provide a lot of information into the sampling procedure, as we observe for θ1 o…
Figure 8
Figure 8. Figure 8: Score estimation loss. “D” indicates debiasing and “C” indicates curvature penalty. [PITH_FULL_IMAGE:figures/full_fig_p070_8.png]
Figure 9
Figure 9. Figure 9: Example of observed data under (a) 5-floor setting and (b) 10-floor setting. [PITH_FULL_IMAGE:figures/full_fig_p072_9.png]
Figure 10
Figure 10. Figure 10: Posterior density plots of different methods in one experiment of 10-floor setting [PITH_FULL_IMAGE:figures/full_fig_p074_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Provable Diffusion Posterior Sampling for Bayesian Inversion

    stat.ML 2025-12 conditional novelty 6.0 of 10

    A diffusion posterior sampler using Monte Carlo Langevin score estimation and warm start is proven to converge in Wasserstein-2 distance under semi-log-concavity and sub-Gaussian assumptions, and outperforms DPS/TV on...

  2. Zeroth-Order Langevin Monte Carlo via SPSA under Noisy Function Measurements

    math.OC 2026-08 reject novelty 5.0 of 10

    LMC-SPSA, using two noisy function evaluations per iteration, is claimed to converge in W2 distance with an O(p^2) dimension bound and O(p/epsilon^2 + delta^2 p^3/epsilon^3) oracle complexity.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages · cited by 2 Pith papers

  1. [2]

    Reference Table:Generate D = {(θ(k), X(k) n )}N k=1 iid ∼ q(θ) p(n) θ (Xn)

    work page

  2. [3]

    Network Training:Train sϕ(θ, X) on D using loss in (23) and obtain bϕ

    work page

  3. [4]

    Return {θ(k)}K k=1 as approximated posterior samples In the actual implementation, we randomly partition the data into a training set (50%) and a validation set (50%)

    Langevin Sampling:For k = 1 to K θ(k) ← θ(k−1) + τn Pn i=1 s bϕ θ(k−1), X∗ i + ∇θ log π(θ(k)) + √2τn Uk, Uk iid ∼ N(0, Idθ). Return {θ(k)}K k=1 as approximated posterior samples In the actual implementation, we randomly partition the data into a training set (50%) and a validation set (50%). We first initialize the neural network with only the score-match...

    work page

  4. [5]

    Localization: Construct a proposal distribution q(θ) using (5)

    work page

  5. [6]

    Reference Table: Generate DS = {(θ(k), X(k))}N k=1 iid ∼ q(θ)pθ(·) and DR = {θ(l), X(l) mR}NR l=1 iid ∼ q(θ)p(mR) θ (·)

    work page

  6. [7]

    Network Training:Train sϕ(θ, X) on DS and DR using loss in (24) and obtain bϕ

    work page

  7. [8]

    Mean Regression:Estimate the mean of s bϕ(θ, X) on DR using (25) and obtain bψ

    work page

  8. [9]

    Debiasing: es(θ, X) = s bϕ(θ, X) − h bψ(θ)

    work page

Show all 12 references
  1. [10]

    Return {θ(k)}K k=1 as approximated posterior samples Ep( · |Xi−1=x,θ)[s∗(θ, x, Xi)] = 0 for each x

    Langevin Sampling:For k = 1 to K θ(k) ← θ(k−1) + τn Pn i=1 ˜s θ(k−1), X∗ i + ∇θ log π(θ(k)) + √2τn Uk, Uk iid ∼ N(0, Idθ). Return {θ(k)}K k=1 as approximated posterior samples Ep( · |Xi−1=x,θ)[s∗(θ, x, Xi)] = 0 for each x. Thus, the two-step debiased score-matching pro- cedure...

  2. [11]

    Alternative 2: This alternative utilizes the idea of projected gradient descent

    Since both A1.2 and A1.3 can be well controlled, it suffices to control A1.1, and we have the following alternative objective to replace (7) as bψ = arg min ψ Eq(θ) hψ(θ) − Epθs bϕ(θ, X) 2 + λ2 hψ(θ)hψ(θ)T + ∇θhψ(θ) 2 F This objective here has a simpler form than (7), although...

  3. [12]

    Minimize the regularized score loss: min ϕ,ψ Eq(θ) Epθ h ∥esϕ,ψ(θ, X) − s∗(θ, X)∥2 i + λ1 Epθ esϕ,ψ(θ, X)esϕ,ψ(θ, X)T + ∇θesϕ,ψ(θ, X) 2 F ,

  4. [13]

    D” indicates debiasing and “C

    Projection: min ψ Eq(θ) hψ(θ) − Epθs bϕ(θ, X) 2 When this procedure converges, we will get a score model within the mean-zero model family that minimizes the curvature regularized score loss. We find Algorithm 3 and the two alternatives have similar empirical performance in ou...

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.