One-step lowest-variance selection in a Gaussian random-field model motivated by masked diffusion: Total correlation and a square root collision threshold

Linjun Li

One-step lowest-variance selection in a Gaussian random-field model motivated by masked diffusion: Total correlation and a square root collision threshold

T0 review · 0 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Within a one-step Gaussian random-field model, the paper establishes a sharp square-root threshold: if the selected budget grows slower than K ~ √N, the dependence cost of the chosen positions vanishes; at the √N scale, it stays bounded awa

desk verdict Clean, internally sound square-root threshold for a stylized Gaussian selection model; the diffusion motivation is honest framing, not a proven bridge. read the letter →

arxiv 2607.17522 v1 pith:X75534YA submitted 2026-07-20 cs.LG cs.AIcs.ITmath.ITmath.PR

Linjun Li This is my paper · ORCID

classification cs.LGcs.AIcs.ITmath.ITmath.PR MSC 60G1560F0562H2094A1760G70

keywords maskeddiffusiontotalcorrelationGaussianrandomfieldsmall-ballprobabilityPoissonapproximationconfidence-basedunmaskingsquare-rootthresholdparalleldecoding

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 isolates a single selection step from masked-diffusion decoding: from a locally dependent nonnegative score field, pick the K positions with smallest scores, then charge the selected positions a distance-dependent Gaussian dependence cost (total correlation). It proves that if K grows like N^α with α < 1/2, this cost vanishes in probability, so a conservative budget can grow without accumulating dependence. At the square-root scale K ~ λ√N, the cost stops vanishing: with positive asymptotic probability it is at least c_ρX, and its expected value is bounded below by c_ρX ϑ λ². The critical mechanism is the collision of selected positions into adjacent pairs, whose count converges to a Poisson law. A reader should care because this identifies √N as an explicit stochastic-geometry baseline for how large a parallel unmasking budget can be before dependence among selected positions becomes unavoidable.

What carries the argument

The central object is the pair (S_N,K, TC_ΣN): the exact set of K smallest values of V_i=‖Y_i‖², where Y is an R^m-valued AR(1) Gaussian chain with correlation ρ_V, and the conditional Gaussian total correlation of an auxiliary AR(1) Gaussian vector with correlation ρ_X on those indices. After standardization, TC_ΣN(S) = −½ log det R_{ρ_X}(S) = Σ_{a=1}^{k-1} −½ log(1−ρ_X^{2(gap_a)}), so the cost is a decreasing function of selected-index spacings. The proof machinery is rare-minima small-ball probability: for distance h, P(both V_i and V_{i+h} below threshold) ~ (1−ρ_V^{2h})^{−m/2} q², which yields a Poisson limit for adjacent selected pairs at K~λ√N and a sparsification theorem (with high p

What would settle it

Take a trained masked-diffusion model, run one confidence-ranked unmasking step at budget K=λ√N over many sequences, and count adjacent selected positions. If the no-adjacent-pair probability does not approach exp(−ϑλ²) (with ϑ estimated from the score field), or if the measured categorical error does not rise at the √N scale, then the model's predicted square-root collision threshold is not what real decoders see.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is a phase transition in the conditional Gaussian total correlation of the exact lowest-variance set in an AR(1) chi-square score-field model. Theorem 2.6: for budgets K_N=⌊N^α⌋ with 0<α<1/2 (more generally when K_N² L_N/N→0 and K_N|ρ_X|^{2(L_N+1)}→0), the cost tends to zero in probability. Theorem 5.1: if K_N/√N→λ∈(0,∞), then liminf P(cost ≥ c_ρX) ≥ 1−exp(−ϑ λ²) > 0 and liminf E[cost] ≥ c_ρX ϑ λ² > 0. The cost is explicitly TC = −½ log det of the AR(1) correlation matrix of the selected indices, equal to a sum over gaps; each adjacent pair contributes exactly c_ρX = −½ log(1−ρ_X²). The proof flows from a Poisson limit for adjacent selected pairs: at threshold

Load-bearing premise

The result assumes the confidence scores cluster like a simple Markov Gaussian field and that the Gaussian dependence cost stands in for real categorical error; if either bridge fails, the square-root warning does not transfer to actual decoders.

Editorial extensions

If this is right

If correct, the square-root scale is the exact critical budget for the stylized cost: sub-√N budgets have vanishing dependence cost, while any fixed positive λ at √N gives a strictly positive lower bound.
The explicit lower bound c_ρX ϑ λ² means the obstruction is quantifiable: the dependence cost is at least the adjacent-pair constant times the Poisson mean.
Supercritical budgets K ≫ √N force the cost to exceed c_ρX with probability tending to 1 (Corollary 5.4).
The Poisson law for adjacent collisions gives a finite-size prediction that can be checked in simulations of the model, and the paper verifies it.
Because the result is invariant under increasing calibrations of the scores, it applies to any monotone transform of the confidence score, not just the raw chi-square values.

Reading between the lines

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

If trained masked models show the same rare-minima clustering of low-confidence scores, √N is a natural ceiling for one-shot parallel unmasking budgets: below it the geometric cost is negligible, at it the cost is dominated by adjacent selected positions.
The model separates ρ_V (score clustering) from ρ_X (decoder dependence). A testable extension: estimate both from real models and compare the observed lower-tail collision rate with the Poisson prediction exp(−ϑ λ²).
The calibration invariance suggests the threshold is not an artifact of the chi-square scale; any monotone confidence measure yields the same selection geometry, so the square-root collision scale should be robust across confidence estimators.
The one-sided nature of the lower bound leaves open whether TC actually converges to a nonzero constant or fluctuates; measuring the full distribution of TC at the critical scale would refine the phase diagram.

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.

Desk Editor's Note

read the letter

Colleague,

What you should know first: this is a genuinely internal result. The paper proves a sharp square-root threshold — sub-root budgets give vanishing conditional Gaussian total correlation, and at K ~ λ√N the cost is bounded below by a positive constant with positive probability — for the lowest-K set of a stationary AR(1) chi-square field. The proof chain is complete: small-ball domination, threshold-set concentration, factorial-moment Poisson limit, and a two-sided sandwich for the exact order statistic. I checked the key estimates in (A.18)–(A.22) and (B.24), and they decay exactly as claimed. This is not a paper where the theorems leak.

What is genuinely new: the square-root collision threshold for the exact lowest-K selection, with explicit constants ϑ and c, plus the total-correlation lower bounds. The machinery itself is standard stochastic geometry, but that is a point in its favor — the result is cleanly isolated.

The soft spot is the bridge, not the math. The score field is chi-square AR(1) and the dependence cost is an independent AR(1) Gaussian kernel. Neither is derived from a trained masked diffusion model. The authors say this clearly in Section 6 and in the remarks around (1.1)–(1.5): they offer a null model, not a transfer theorem. I do not see this as a flaw in the paper's own claims, just a hard limit on what the paper can tell you about actual diffusion. The synthetic checks in Section 3.2 simulate the same model, so they confirm the asymptotics but not the motivation. That is fine as long as it is tagged, and it is tagged.

One minor thing: Theorem 5.1 gives one-sided lower bounds, not a phase transition for total correlation itself. The authors explicitly disclaim a complete picture, so it is not overclaiming, but the abstract's 'square-root collision threshold' may read as a stronger statement than the adjacent-pair Poisson limit supports.

Who this is for: anyone working on confidence-based unmasking schedules who wants a rigorous baseline for how rare-minima clustering can create dependence costs. It is also a useful example of a clean stochastic-geometry analysis of an order statistic. I would send it to a serious referee. The internal results deserve to be on record, and the limitations are stated rather than hidden.

Referee Report

0 major / 4 minor

Summary. The paper analyzes a single static selection step in a stylized Gaussian random-field model. An R^m Gaussian AR(1) chain (Y_i) yields a nonnegative score field V_i=||Y_i||^2; the scheduler selects the K smallest scores. A separate AR(1) Gaussian kernel with parameter ρ_X assigns a conditional Gaussian total-correlation cost TC_{Σ_N}(S)=−1/2 log det R_{ρ_X}(S), so the cost depends only on the spacings of the selected indices. The first main theorem (Thm 2.6) shows that TC_{Σ_N}(S_{N,K}) → 0 in probability if L_N K_N^2 / N → 0 and K_N |ρ_X|^{2(L_N+1)} → 0; this covers polynomial budgets K_N = N^α with α < 1/2. The second main theorem (Thm 5.1) shows that if K_N / √N → λ ∈ (0,∞), then liminf P(TC ≥ c_{ρ_X}) ≥ 1 − exp(−ϑ_{ρ_V,m} λ^2) > 0 and liminf E[TC] ≥ c_{ρ_X} ϑ_{ρ_V,m} λ^2 > 0. The proofs use an auxiliary sparsification theorem (Thm 4.1) and a Poisson limit for adjacent selected pairs (Thm 5.2). The paper explicitly frames the Gaussian model as a stylized baseline, not as a derived model of a trained masked-diffusion decoder.

Significance. This is a self-contained stochastic-geometry contribution. Within the stated one-step Gaussian model, the square-root budget is exactly the critical scale at which adjacent rare minima create a non-negligible dependence cost; the proof chain in Appendices A–B is complete and internally consistent. The paper is unusually transparent about its scope: §6 and the remarks around (1.1)–(1.5) explicitly disclaim that the Gaussian score field, the AR(1) dependence kernel, or Gaussian total correlation are derived from a categorical decoder, and the synthetic checks in §3.2 are internal consistency checks of the same model rather than external validation. The strength of the paper is therefore a rigorous null model for one-step confidence-based selection, with the transfer to actual masked diffusion left as a concrete empirical question. I find no load-bearing technical defect; the external-validity gap is real but acknowledged and does not undermine the internal theorems.

minor comments (4)

[§2.1, after Prop. 2.1] Typo: 'the pair (Vi,RhoX,N)' should read 'the pair (V_i, R_{\rho_X,N})' or simply '(V_i, \Sigma_N)'.
[§B.3, Prop. B.3] The sentence 'ordinary moments are finite linear combinations of factorial moments of lower order' is imprecise; ordinary moments are combinations of factorial moments of order at most r. The uniform-integrability step is also terse; a one-sentence justification would improve readability.
[§B.2, Lemma B.2] The statement that the o(1) term 'may depend on (r,ρV,m,qN,gN)' is awkward: the rate may depend on the sequences, but the uniformity is in block locations. Please rephrase.
[§5, Remark after Cor. 5.4] The scope remark correctly notes that no complete phase diagram for the total-correlation statistic is claimed. It would help to state this explicitly in the introduction to §5, so the one-sided lower bound is not over-read.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theorems are self-contained derivations from explicit model assumptions, with no fitted parameters and no load-bearing self-citations.

full rationale

The paper’s central claims are internal to an explicitly stated one-step Gaussian random-field model. The derivation chain is: Lemma 2.4 gives the exact AR(1) determinant; Lemma 2.5 bounds total correlation for separated sets; Theorem 4.1 proves sparsification of exact lowest-K selection from the chi-square small-ball estimates in Lemmas A.1–A.3 and the threshold-enlargement comparison Lemma A.4; Theorem 2.6 then follows deterministically from Theorem 4.1 plus Lemma 2.5. The complementary result Theorem 5.2 is proved by a Poisson factorial-moment argument (Proposition B.3) and a two-sided threshold sandwich; Theorem 5.1 then follows from the exact identity TC(S) ≥ c_rhoX A(S) (Eq. 5.6). No parameter is fitted to data, no constant is set by the quantity being predicted, and no ‘uniqueness theorem’ or ansatz is imported from prior work by the same author. The only self-referential element is the synthetic check in Section 3.2, which simulates the same model; the paper explicitly labels this as checking ‘the internal scaling of the model, not a trained masked-diffusion decoder’ and does not use it to justify the theorems. The limitations in Section 6 concede that the result does not transfer to trained masked models unless additional empirical properties hold; this is a scope limitation, not a circular derivation. Therefore no circular step is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The paper is honest about its three invented inputs: the chi-square score field, the AR(1) score dependence, and the AR(1) Gaussian dependence kernel. The theorems are universal over the allowed parameter ranges, so none of these inputs is fitted; they are model postulates. The proof itself adds no new free constants beyond the proof-device threshold densities q_N.

free parameters (3)

m (chi-square degrees of freedom) = not fitted (m=3 in simulations)
Defines the score distribution V_i ∼ χ²_m; theorems hold for fixed m; chosen by hand for experiments.
ρ_V (score-field AR(1) correlation) = 0.5 in simulations
Controls lower-tail clustering of the score field; external model input, not estimated from data.
ρ_X (dependence-kernel AR(1) correlation) = 0.5 in simulations
Sets the Gaussian total-correlation cost decay; external model input, not estimated from data.

assumptions (4)

ad hoc to paper Scores V_i = ||Y_i||^2 with (Y_i) a stationary R^m Gaussian AR(1) chain (Eq. 2.1).
Defines the model; no evidence trained confidence scores follow this field.
ad hoc to paper Gaussian cost: conditional covariance Σ_N = D_V^{1/2} R_{ρ_X,N} D_V^{1/2} and TC = −1/2 log det R(S) (Eqs. 1.3, 2.20–2.21).
The dependence cost is prescribed, not derived from a categorical decoder.
domain assumption Selection rule: choose the exact K smallest V_i; ties have probability zero (Section 2.3).
Scheduler rule; within the model.
standard math Standard Gaussian small-ball estimates, factorial moments, and Poisson approximation facts.
Used throughout Appendices A and B.

invented entities (2)

Gaussian score field (Y_i, V_i)
purpose: Represent positionwise uncertainty and define the lowest-score set.
A stylized substitute for neural confidence scores; no calibration to any trained model.
Auxiliary Gaussian vector X^{(N)} with random covariance Σ_N
purpose: Define the conditional Gaussian total-correlation cost on the selected indices.
Analytically convenient cost; explicitly not derived from categorical decoder.

how reviews work

0 comments

Cite this review

Pith. "Pith review of One-step lowest-variance selection in a Gaussian random-field model motivated by masked diffusion: Total correlation and a square root collision threshold." pith.science (2026). https://pith.science/paper/X75534YA

@misc{pith2026260717522,
  author       = {Pith},
  title        = {Pith review of: One-step lowest-variance selection in a Gaussian random-field model motivated by masked diffusion: Total correlation and a square root collision threshold},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X75534YA}},
  note         = {Machine review of arXiv:2607.17522}
}

read the original abstract

Motivated by confidence-guided parallel unmasking in masked discrete diffusion, we study a single selection step in a stylized Gaussian random-field model. A locally dependent nonnegative score field represents position wise uncertainty, and the scheduler selects the K positions with the smallest scores. Dependence among the selected positions is measured through a distance-dependent Gaussian correlation model. This separation provides a tractable framework for quantifying how the geometry of low-score locations affects the dependence cost of factorized parallel decoding. We establish two complementary results. In a conservative sub-square-root regime, the conditional Gaussian total correlation of the selected block vanishes in probability. At the square-root scale, it remains non-negligible with positive asymptotic probability and admits a strictly positive expectation lower bound. Synthetic experiments support the predicted finite-size behavior. These results provide a rigorous stochastic-geometry baseline for understanding how budget size, score dependence, and spatial correlation jointly shape one-step confidence-based selection in masked discrete diffusion.

Figures

Figures reproduced from arXiv: 2607.17522 by the authors.

**Figure 2.** Probability of no adjacent selected pair in the stylized score model for [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 1 linked inside Pith

[1]

Structured denoising diffusion models in discrete state-spaces

Austin, J., Johnson, D.D., Ho, J., Tarlow, D., van den Berg, R., 2021. Structured denoising diffusion models in discrete state-spaces. Adv. Neural Inf. Process. Syst. 34, 17981--17993

2021
[2]

A continuous time framework for discrete denoising models

Campbell, A., Benton, J., De Bortoli, V., Rainforth, T., Deligiannidis, G., Doucet, A., 2022. A continuous time framework for discrete denoising models. Adv. Neural Inf. Process. Syst. 35, 28266--28279

2022
[3]

MaskGIT: Masked generative image transformer, in: Proc

Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T., 2022. MaskGIT: Masked generative image transformer, in: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 11315--11325

2022
[4]

Muse: Text-to-image generation via masked generative transformers, in: Proc

Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.-H., Murphy, K.P., Freeman, W.T., Rubinstein, M., Li, Y., Krishnan, D., 2023. Muse: Text-to-image generation via masked generative transformers, in: Proc. 40th Int. Conf. Mach. Learn., Proc. Mach. Learn. Res. 202, 4055--4075

2023
[5]

Optimal inference schedules for masked diffusion models, in: Proc

Chen, S., Cong, K., Li, J., 2026. Optimal inference schedules for masked diffusion models, in: Proc. Thirty-Ninth Conf. Learn. Theory, Proc. Mach. Learn. Res. 336, 1279--1311

2026
[6]

Elements of Information Theory, second ed

Cover, T.M., Thomas, J.A., 2006. Elements of Information Theory, second ed. Wiley, Hoboken. https://doi.org/10.1002/047174882X

doi:10.1002/047174882x 2006
[7]

Improving discrete diffusion unmasking policies beyond explicit reference policies, in: Fourteenth International Conference on Learning Representations; arXiv:2510.05725

Hong, C., An, S., Kim, M.-S., Ye, J.C., 2026. Improving discrete diffusion unmasking policies beyond explicit reference policies, in: Fourteenth International Conference on Learning Representations; arXiv:2510.05725

arXiv 2026
[8]

Argmax flows and multinomial diffusion: Learning categorical distributions

Hoogeboom, E., Nielsen, D., Jaini, P., Forr\'e, P., Welling, M., 2021. Argmax flows and multinomial diffusion: Learning categorical distributions. Adv. Neural Inf. Process. Syst. 34, 12454--12465

2021

Show all 18 references

[9]

Autoregressive diffusion models, in: International Conference on Learning Representations

Hoogeboom, E., Gritsenko, A.A., Bastings, J., Poole, B., van den Berg, R., Salimans, T., 2022. Autoregressive diffusion models, in: International Conference on Learning Representations

2022
[10]

Learning unmasking policies for diffusion language models, in: Forty-Third International Conference on Machine Learning (oral spotlight)

Jazbec, M., Olausson, T.X., B\'ethune, L., Ablin, P., Kirchhof, M., Monteiro, J., Turrisi, V., Ramapuram, J., Cuturi, M., 2026. Learning unmasking policies for diffusion language models, in: Forty-Third International Conference on Machine Learning (oral spotlight). arXiv:2512.09106

2026 arXiv
[11]

Train for the worst, plan for the best: Understanding token ordering in masked diffusions, in: Proc

Kim, J., Shah, K., Kontonis, V., Kakade, S.M., Chen, S., 2025. Train for the worst, plan for the best: Understanding token ordering in masked diffusions, in: Proc. 42nd Int. Conf. Mach. Learn., Proc. Mach. Learn. Res. 267, 30749--30768

2025
[12]

Breaking AR's sampling bottleneck: Provable acceleration via diffusion language models

Li, G., Cai, C., 2025. Breaking AR's sampling bottleneck: Provable acceleration via diffusion language models. Adv. Neural Inf. Process. Syst. 38

2025
[13]

Discrete diffusion modeling by estimating the ratios of the data distribution, in: Proc

Lou, A., Meng, C., Ermon, S., 2024. Discrete diffusion modeling by estimating the ratios of the data distribution, in: Proc. 41st Int. Conf. Mach. Learn., Proc. Mach. Learn. Res. 235, 32819--32848

2024
[14]

Large language diffusion models

Nie, S., Zhu, F., You, Z., Zhang, X., Ou, J., Hu, J., Zhou, J., Lin, Y., Wen, J.-R., Li, C., 2025. Large language diffusion models. Adv. Neural Inf. Process. Syst. 38

2025
[15]

Simple and effective masked diffusion language models

Sahoo, S.S., Arriola, M., Schiff, Y., Gokaslan, A., Marroquin, E., Chiu, J.T., Rush, A., Kuleshov, V., 2024. Simple and effective masked diffusion language models. Adv. Neural Inf. Process. Syst. 37, 130136--130184. https://doi.org/10.52202/079017-4135

2024 doi
[16]

The diffusion duality, in: Proc

Sahoo, S.S., Deschenaux, J., Gokaslan, A., Wang, G., Chiu, J.T., Kuleshov, V., 2025. The diffusion duality, in: Proc. 42nd Int. Conf. Mach. Learn., Proc. Mach. Learn. Res. 267, 52584--52619

2025
[17]

The unreasonable effectiveness of Gaussian score approximation for diffusion models and its applications

Wang, B., Vastola, J.J., 2024. The unreasonable effectiveness of Gaussian score approximation for diffusion models and its applications. Trans. Mach. Learn. Res

2024
[18]

Information theoretical analysis of multivariate correlation

Watanabe, S., 1960. Information theoretical analysis of multivariate correlation. IBM J. Res. Dev. 4, 66--82. https://doi.org/10.1147/rd.41.0066

1960 doi

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Structured denoising diffusion models in discrete state-spaces

Austin, J., Johnson, D.D., Ho, J., Tarlow, D., van den Berg, R., 2021. Structured denoising diffusion models in discrete state-spaces. Adv. Neural Inf. Process. Syst. 34, 17981--17993

2021

[2] [2]

A continuous time framework for discrete denoising models

Campbell, A., Benton, J., De Bortoli, V., Rainforth, T., Deligiannidis, G., Doucet, A., 2022. A continuous time framework for discrete denoising models. Adv. Neural Inf. Process. Syst. 35, 28266--28279

2022

[3] [3]

MaskGIT: Masked generative image transformer, in: Proc

Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T., 2022. MaskGIT: Masked generative image transformer, in: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 11315--11325

2022

[4] [4]

Muse: Text-to-image generation via masked generative transformers, in: Proc

Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.-H., Murphy, K.P., Freeman, W.T., Rubinstein, M., Li, Y., Krishnan, D., 2023. Muse: Text-to-image generation via masked generative transformers, in: Proc. 40th Int. Conf. Mach. Learn., Proc. Mach. Learn. Res. 202, 4055--4075

2023

[5] [5]

Optimal inference schedules for masked diffusion models, in: Proc

Chen, S., Cong, K., Li, J., 2026. Optimal inference schedules for masked diffusion models, in: Proc. Thirty-Ninth Conf. Learn. Theory, Proc. Mach. Learn. Res. 336, 1279--1311

2026

[6] [6]

Elements of Information Theory, second ed

Cover, T.M., Thomas, J.A., 2006. Elements of Information Theory, second ed. Wiley, Hoboken. https://doi.org/10.1002/047174882X

doi:10.1002/047174882x 2006

[7] [7]

Improving discrete diffusion unmasking policies beyond explicit reference policies, in: Fourteenth International Conference on Learning Representations; arXiv:2510.05725

Hong, C., An, S., Kim, M.-S., Ye, J.C., 2026. Improving discrete diffusion unmasking policies beyond explicit reference policies, in: Fourteenth International Conference on Learning Representations; arXiv:2510.05725

arXiv 2026

[8] [8]

Argmax flows and multinomial diffusion: Learning categorical distributions

Hoogeboom, E., Nielsen, D., Jaini, P., Forr\'e, P., Welling, M., 2021. Argmax flows and multinomial diffusion: Learning categorical distributions. Adv. Neural Inf. Process. Syst. 34, 12454--12465

2021