A Kinematic Constraint on Pedestrian Walking: Power-law Scaling between Critical Angular Velocity and Speed

Chao Li; Jinghui Wang; Wei Lv; Yufei Li

REVIEW 4 major objections 5 minor 23 references

A Kinematic Constraint on Pedestrian Walking: Power-law Scaling between Critical Angular Velocity and Speed

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Pedestrian turning rate has a speed limit: the fastest allowable turn falls as walking speed increases.

desk verdict The paper's universal turning-limit exponent is likely an artifact of the curvature estimator's noise floor, but the question and data are worth a serious revision. read the letter →

arxiv 2506.15321 v1 pith:H6HT6IG4 submitted 2025-06-18 physics.soc-ph

Jinghui Wang , Wei Lv , Chao Li , Yufei Li This is my paper · ORCID

classification physics.soc-ph

keywords pedestriandynamicsangularvelocitycurvaturepower-lawscalingkinematicconstraintcrowdtrajectoriesturningbehaviorphasespace

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 claims that pedestrian turning is governed by a kinematic speed limit: the highest angular velocity a walker can sustain falls as $\omega_{\mathrm{crit}} \propto v^{-0.8}$, equivalently critical curvature falls as $\kappa_{\mathrm{crit}} \propto v^{-1.8}$. The claim is based on re-analyzing trajectories from nine controlled experiments covering crossing flows, unidirectional and bidirectional corridors, pedestrian-vehicle interaction, single-file walking, and antipodal circle crossings. In every dataset and at two sampling rates, the upper edge of the speed-angular velocity scatter cloud follows the same power law, so the authors interpret it as a universal boundary separating kinematically possible from impossible pedestrian motion. A reader should care because, if true, this is a simple quantitative constraint that any model of pedestrian motion, from crowd simulations to trajectory prediction, would need to respect.

What carries the argument

The load-bearing object is the critical envelope of the scatter plot of instantaneous speed versus angular velocity, computed from discrete trajectories as $\kappa = d\theta/d\ell \simeq \alpha/s = \omega/v$, where $\alpha$ is the heading change in a sampling interval and $s$ the distance traveled. On log-log axes the upper edge of each cloud is fitted, by a subjectively drawn trend curve, to a power law with a fixed exponent, giving $\kappa_{\mathrm{crit}} \propto \pm v^{-1.8}$ and hence $\omega_{\mathrm{crit}} \propto \pm v^{-0.8}$ (Eq. 3). This envelope, not any individual trajectory, is what carries the argument that a kinematic turning constraint exists and that it scales with speed.

What would settle it

Refit the upper envelope after randomly subsampling trajectories from each of the nine datasets; if the fitted exponent changes systematically with sample size, or if surrogate trajectories built from the same speed distribution and noise but with unconstrained random headings show the same $-0.8$ power-law envelope, then the claimed kinematic constraint is a statistical artifact rather than a physical boundary.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that pedestrian trajectories occupy a bounded region of the $(\omega, v)$ phase space whose upper envelope obeys a power law with exponent approximately $-0.8$, rather than a fixed angular-velocity ceiling. Because curvature is defined as $\kappa = \omega/v$, this is equivalent to a critical-curvature law $\kappa_{\mathrm{crit}} \propto v^{-1.8}$. The bound appears consistently across nine datasets with different geometries, densities, interaction types, filtering schemes, and sampling intervals, which the authors take as evidence that it is a kinematic constraint intrinsic to walking rather than a feature of any single experiment or measurement setup.

Load-bearing premise

The load-bearing premise is that the upper edge of the scatter cloud, which the authors themselves describe as a subjectively defined trend curve, is a genuine kinematic boundary rather than an artifact created by finite sample sizes, measurement noise, and the way maxima grow with the number of trajectories sampled.

Editorial extensions

If this is right

At higher walking speeds, sharp turns become kinematically less available: the maximum sustainable angular velocity decreases, so rapid direction changes must be spread over longer distances or preceded by deceleration.
The same bound implies a minimum turning radius that grows with speed: because $\kappa \propto v^{-1.8}$, the radius of the tightest achievable turn scales roughly as $v^{1.8}$.
Trajectory prediction and crowd-simulation models that ignore the bound will over-generate high-speed, high-curvature motion that real pedestrians do not produce.
The speed-angular velocity phase space, not just the classical speed-density fundamental diagram, should be treated as a basic descriptor of pedestrian motion.
Because the law is reported at both frame-by-frame and decimated sampling, the constraint is claimed to be stable with respect to temporal resolution.

Reading between the lines

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

The paper does not derive the exponent; a biomechanical model based on step frequency and stride length might explain $-0.8$ and could be tested against these same datasets.
The paper pools all trajectories into one cloud; a stricter test would fit per-pedestrian critical curves and ask whether the exponent survives within a single walker.
If the bound is truly kinematic, it should be reproducible under forced circular walking at controlled speeds, with the tightest sustainable radius growing as $v^{1.8}$.
A surrogate-data check would separate physics from statistics: random-walk headings with identical speed and noise statistics should not reproduce the $-0.8$ envelope if the constraint is real.

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.

Desk Editor's Note

read the letter

Colleague,

The headline claim here — a universal ωcrit ∝ v^{-0.8} bound across nine pedestrian datasets — is not supported by the evidence as presented. The exponent is imposed, not measured. Section 3.2 states the trend curve is 'subjectively defined' with a 'fixed exponent of -1.8' in curvature; the ω exponent is then just the algebraic restatement (ω = κv). No fit statistics, confidence intervals, or null models appear anywhere. The stress-test worry lands hard: for straight walking with fixed tracking noise, the estimator in Eq. 2 gives κ errors that scale as v^{-2} and ω errors as v^{-1}, so the largest estimated values in a speed bin decline with speed even in the complete absence of a kinematic constraint. The reported exponents (-1.8, -0.8) are close enough to those null scalings that the envelope may be nothing more than the noise floor of the estimator.

What the paper does well: it assembles a genuinely broad set of independent trajectory datasets, applies a variety of preprocessing filters, and examines two sampling rates. That is real work, and the data are open. The question — does a speed-dependent turning bound exist? — is a good one. The bounded-region picture is plausible and consistent with the known speed-curvature trade-off in the literature.

The soft spots are not minor; they are load-bearing. The envelope is fit by eye. The exponent is fixed a priori, which is circular when the same scatter is said to confirm it. The decimation analysis shows changes in the constant coefficient but no robustness check on the exponent. And the paper does not report any negative control, such as computing the same envelopes from straight-line walking.

Bottom line: this is a paper that deserves a serious referee because the question and data are worth engaging, but it is not publishable in its current form. A revision would need a proper envelope estimator (e.g., quantile regression or extreme-value methods), a null-model test on straight-walk data, and honest error bars on the exponent. If the exponent survives those tests, it would be a real result. I would not cite it yet, but I would bring it to a reading group as a case study in estimator noise masquerading as a scaling law.

Referee Report

4 major / 5 minor

Summary. The paper analyzes pedestrian trajectories from nine experiments, computes curvature and speed from position samples via Eq. (2), and plots scatter diagrams. The authors draw a "subjectively defined" trend curve with a fixed exponent of -1.8 for the upper envelope of curvature versus speed, and then state that this implies a power-law scaling between critical angular velocity and speed, ωcrit ∝ v^{-0.8} (Eq. 3). They interpret this as a universal kinematic constraint bounding the speed–angular-velocity phase space of pedestrian motion. The manuscript provides no quantitative fitting procedure, no goodness-of-fit measures, and no control for measurement noise.

Significance. If the claimed scaling were established, it would offer a simple kinematic constraint useful for pedestrian modeling and trajectory prediction. The paper compiles a large set of heterogeneous experimental data, which is a strength. However, the central evidence is an eyeballed envelope with a fixed exponent, and the angular-velocity scaling is an algebraic restatement of the curvature scaling. Moreover, the reported exponents are close to those expected from the estimator's noise floor for straight walking, and the paper provides no null test. As it stands, the central claim is not supported by the presented analysis.

major comments (4)

[Section 3.2, Eq. (3)] The trend curve is described as "subjectively defined" with a "fixed exponent of -1.8" in curvature. No fitting procedure, confidence intervals, goodness-of-fit statistics, or comparison with alternative exponents is reported. The exponent is therefore imposed rather than estimated from the data, and this is the sole evidence for the claimed universal power law.
[Eq. (3)] Because Eq. (2) defines ω = κ v, the statement ωcrit ∝ v^{-0.8} is a direct algebraic rearrangement of κcrit ∝ v^{-1.8}. The angular-velocity scaling is not an independent empirical finding; the paper's headline claim about ω carries no information beyond the κ–v envelope.
[Section 3.2, Figs. 12–20] The upper envelope of the scatter clouds is interpreted as a kinematic boundary, but no test rules out the null hypothesis that it arises from the estimator's noise floor. For straight-line walking with perpendicular tracking noise ε, the error in curvature estimated by Eq. (2) is of order ε/(v^2 Δt^2), so the largest estimated κ in a speed bin scales as v^{-2} and the largest estimated ω = κ v scales as v^{-1}. The reported exponents, -1.8 and -0.8, are close to these null scalings, yet no straight-walk control, synthetic-noise test, or sample-size analysis is provided to show that the envelope is a genuine kinematic constraint.
[Section 3.2, sampling configuration] The paper does not report how many samples fall in each speed bin or whether the upper envelope is stable under subsampling. Since the maximum value of a noisy sample cloud increases with sample size, a declining envelope can appear simply from heterogeneous sample counts across speed bins, even with no physical turning limit. This alternative explanation is not addressed.

minor comments (5)

[Section 2.1] In the paragraph on trajectory errors, "will introducing" should be "will introduce."
[Table 1] The Sampling Interval column lists two values for each dataset; the caption should clarify which value corresponds to the frame-by-frame sampling and which to the decimated sampling, and how these relate to the Frame Rate column.
[Eq. (3)] The use of the plus-minus sign in κcrit ∝ ±v^{-1.8} is confusing; the scaling presumably applies to the absolute value of curvature, and this should be stated explicitly.
[Figures 12–20] The captions do not identify the axes or their units clearly. The figures should state explicitly that the vertical axis is curvature (or angular velocity) and the horizontal axis is speed, with units.
[Data availability] The data are hosted on a Google Drive folder without a DOI or permanent identifier; a stable repository DOI would improve reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline ω∝v^{−0.8} relation is the algebraic restatement of the subjectively fitted κ∝v^{−1.8} envelope (Eqs. 2–3), so the central 'prediction' reduces to a fit on the same data.

self definitional [Section 3.2, Eq. (3), with Eq. (2) as the defining identity]
"κcrit∝±v−1.8 ≜ωcrit∝±v−0.8 (3)"

Equation 2 defines κ = ω/v, so substituting the fitted form κcrit ∝ v^{−1.8} into ω = κv gives ωcrit ∝ v^{−0.8} identically. The angular-velocity scaling claim is therefore not an independently measured result; it is the curvature fit rewritten through the defining relation ω = κv, and no separate angular-velocity data analysis is reported.
fitted input called prediction [Section 3.2]
"For each figure, a subjectively defined trend curve and its corresponding function were established. The scatter data distribution exhibits a clear and consistent trend that closely follows the defined power-law curve with a fixed exponent of -1.8."

The 'critical' curve and its exponent are placed by hand on the same scatter plots that are then said to 'closely follow' it. Because the exponent is fixed in advance ('with a fixed exponent of -1.8') rather than estimated with uncertainty, the claimed confirmation is a restatement of the drawn envelope; no held-out data, null model, goodness-of-fit, or error analysis is used to test it.

full rationale

The central result is a bounded speed–angular velocity region claimed to follow Eq. 3. That claim reduces to two steps: (i) a subjectively placed upper envelope with a fixed exponent −1.8 on the κ–v scatter, and (ii) an algebraic conversion to ω via Eq. 2, κ = ω/v. Neither step is an independent prediction: the ω ∝ v^{−0.8} statement is Eq. 2 applied to the κ fit, and the κ fit itself is confirmed only against the same data used to draw it. The paper supplies no null model, goodness-of-fit statistic, or validation on a separate subset, so the headline scaling is a fitted description rather than a tested constraint. The breadth of nine datasets reduces the risk that the pattern appears in only one sample, but it does not break the circularity because every dataset is treated with the same subjective curve placement. No load-bearing self-citation was found: Experiments 1 and 7 were collected by the authors, but the datasets are external, and no theoretical uniqueness or prior result by the same authors is invoked to force the power law. Finite-sample maxima and tracker-noise floors could also produce a declining envelope in the absence of a kinematic bound; that is a substantive correctness risk, but it is secondary to the circularity already present in Eqs. 2–3 and Section 3.2.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a freely chosen exponent and a subjective envelope placement, plus the assumption that the measured upper edge reflects kinematics rather than sampling noise. No mechanistic derivation is offered, and no independent test separates the envelope from artifacts.

free parameters (3)

power-law exponent for critical curvature (κcrit) = -1.8 (equivalently -0.8 for ωcrit)
The exponent is fixed in the trend curves of Section 3.2, but no fitting procedure or uncertainty is reported; it is a free parameter chosen to match the envelope.
envelope curve constant/offset per dataset and sampling interval = not reported
The paper says scale effects cause 'linear expansion or contraction of the trend lines', meaning per-dataset constants vary, but no values, fitting method, or confidence intervals are given.
filtering parameters (mean filter window, NLMS settings) = mean filter k=5; NLMS settings unspecified
Preprocessing choices affect angular velocity estimates; NLMS parameters are not reported (Section 2.1), so the influence on the envelope is uncontrolled.

assumptions (3)

domain assumption The finite-difference quotient α/s accurately approximates true curvature at the adopted sampling rates.
Equation 2 uses κ ≈ α/s; if trajectories are undersampled or noisy, curvature estimates, especially maxima, are biased.
domain assumption The tracked head position, after filtering, represents the pedestrian's actual motion.
Section 2.1 acknowledges head oscillation and applies filtering; residual error would inflate angular velocity and could create the envelope.
domain assumption The upper envelope of observed points is the critical kinematic limit of pedestrian motion.
Section 3.2 treats the cloud boundary as a physical constraint; observed maxima depend on sample size and noise, so the boundary may be a statistical artifact.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Kinematic Constraint on Pedestrian Walking: Power-law Scaling between Critical Angular Velocity and Speed." pith.science (2026). https://pith.science/paper/H6HT6IG4

@misc{pith2026250615321,
  author       = {Pith},
  title        = {Pith review of: A Kinematic Constraint on Pedestrian Walking: Power-law Scaling between Critical Angular Velocity and Speed},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H6HT6IG4}},
  note         = {Machine review of arXiv:2506.15321}
}

read the original abstract

This paper presents a statistical analysis of speed and angular velocity obtained from pedestrian experiments across nine distinct datasets. Experimental scenarios included crossing motion, unidirectional/bidirectional flows, bidirectional/four-directional crossing flows, pedestrian-vehicle interactions, unidirectional flow in a circular corridor, and circle antipode configurations. We applied filtering methods to reduce noise and analyzed the data at different sampling frequencies. The results reveal a universal power-law scaling between critical angular velocity and speed, with a scaling exponent of approximately -0.8. This relationship defines a bounded region in the speed-angular velocity phase space, suggesting a kinematic constraint on pedestrian motion.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 20 canonical work pages

[1]

A., Bacik, B

Bacik, K. A., Bacik, B. S., & Rogers, T. 2023, Science, 379, 923, doi: 10.1126/science.add8091

doi:10.1126/science.add8091 2023
[2]

2017, Journal of Statistical Mechanics: Theory and Experiment, 2017, 033404, doi: 10.1088/1742-5468/aa620d

Cao, S., Seyfried, A., Zhang, J., Holl, S., & Song, W. 2017, Journal of Statistical Mechanics: Theory and Experiment, 2017, 033404, doi: 10.1088/1742-5468/aa620d

work page doi:10.1088/1742-5468/aa620d 2017
[3]

2016, Physical Review E, 94, 012312, doi: 10.1103/PhysRevE.94.012312

Cao, S., Zhang, J., Salden, D., et al. 2016, Physical Review E, 94, 012312, doi: 10.1103/PhysRevE.94.012312

work page doi:10.1103/physreve.94.012312 2016
[4]

2023, Crowd

Cordes, J., Chraibi, M., & Tordeux, A. 2023, Crowd

work page 2023
[5]

Dynamics, Volume 4: Analytics and Human Factors in Crowd Modeling, 143, doi: 10.1007/978-3-031-46359-4_6

work page doi:10.1007/978-3-031-46359-4_6
[7]

K., & Wood, D

Elliott, J., Simms, C. K., & Wood, D. P. 2012, Accident Analysis & Prevention, 45, 342, doi: 10.1016/j.aap.2011.07.022

work page doi:10.1016/j.aap.2011.07.022 2012
[8]

2017, PloS one, 12, e0169734, doi: 10.1371/journal.pone.0169734

Prattichizzo, D. 2017, PloS one, 12, e0169734, doi: 10.1371/journal.pone.0169734

doi:10.1371/journal.pone.0169734 2017
[9]

2016, Physical Review E, 94, 032304, doi: 10.1103/PhysRevE.94.032304

Feliciani, C., & Nishinari, K. 2016, Physical Review E, 94, 032304, doi: 10.1103/PhysRevE.94.032304

work page doi:10.1103/physreve.94.032304 2016

Show all 23 references

[10]

2019, Physical Review E, 99, 062307, doi: 10.1103/PhysRevE.99.062307

Fujita, A., Feliciani, C., Yanagisawa, D., & Nishinari, K. 2019, Physical Review E, 99, 062307, doi: 10.1103/PhysRevE.99.062307

2019 doi
[11]

Miller, H. H. 1935, in Highway Research Board

1935
[12]

2018, in Proceedings of the IEEE conference on computer vision and pattern recognition, 2255–2264

Gupta, A., Johnson, J., Fei-Fei, L., Savarese, S., & Alahi, A. 2018, in Proceedings of the IEEE conference on computer vision and pattern recognition, 2255–2264

2018
[13]

2005, Experimental brain research, 162, 145, doi: 10.1007/s00221-004-2122-8

Berthoz, A. 2005, Experimental brain research, 162, 145, doi: 10.1007/s00221-004-2122-8

2005 doi
[14]

2011, in Pedestrian and Evacuation Dynamics (Springer), 255–264, doi: 10.1007/978-1-4419-9725-8_23

Hoogendoorn, S., Campanella, M., & Daamen, W. 2011, in Pedestrian and Evacuation Dynamics (Springer), 255–264, doi: 10.1007/978-1-4419-9725-8_23

2011 doi
[15]

2014, PloS one, 9, e89589, doi: 10.1371/journal.pone.0089589

Huber, M., Su, Y.-H., Krüger, M., et al. 2014, PloS one, 9, e89589, doi: 10.1371/journal.pone.0089589

2014 doi
[16]

2020, Journal of the Royal Society Interface, 17, 20200396, doi: 10.1098/rsif.2020.0396 MathWorks

Kleinmeier, B., Köster, G., & Drury, J. 2020, Journal of the Royal Society Interface, 17, 20200396, doi: 10.1098/rsif.2020.0396 MathWorks. 2023, Enhance a Signal Using LMS and Normalized LMS Algorithms. https://www.mathworks.com/help/dsp/ug/ enhance-a-signal-using-lms-and-norm...

2020
[17]

2019, Scientific reports, 9, 105, doi: 10.1038/s41598-018-36711-7

Appert-Rolland, C. 2019, Scientific reports, 9, 105, doi: 10.1038/s41598-018-36711-7

2019 doi
[18]

R., Negri, P

Parisi, D. R., Negri, P. A., & Bruno, L. 2016, Physical Review E, 94, 022318, doi: 10.1103/PhysRevE.94.022318

2016 doi
[19]

Saberi, M., & Mahmassani, H. S. 2014, Transportation research record, 2421, 31, doi: 10.3141/2421-04

2014 doi
[20]

2005, Journal of Statistical Mechanics: Theory and Experiment, 2005, P10002, doi: 10.1088/1742-5468/2005/10/P10002

Seyfried, A., Steffen, B., Klingsch, W., & Boltes, M. 2005, Journal of Statistical Mechanics: Theory and Experiment, 2005, P10002, doi: 10.1088/1742-5468/2005/10/P10002

2005 doi
[21]

2023, arXiv preprint arXiv:2311.04827, doi: 10.48550/arXiv.2311.04827

Wang, J., Lv, W., Jiang, H., Fang, Z., & Ma, J. 2023, arXiv preprint arXiv:2311.04827, doi: 10.48550/arXiv.2311.04827

2023 doi arXiv
[22]

2019, Transportation Research Part C: Emerging Technologies, 103, 174, doi: 10.1016/j.trc.2019.04.007

Xiao, Y., Gao, Z., Jiang, R., et al. 2019, Transportation Research Part C: Emerging Technologies, 103, 174, doi: 10.1016/j.trc.2019.04.007

2019 doi
[23]

2019, in 2019 IEEE Intelligent Vehicles Symposium (IV), IEEE, 899–904, doi: 10.1109/IVS.2019.8814092

Yang, D., Li, L., Redmill, K., & Özgüner, Ü. 2019, in 2019 IEEE Intelligent Vehicles Symposium (IV), IEEE, 899–904, doi: 10.1109/IVS.2019.8814092

2019
[24]

2012, Journal of Statistical Mechanics: Theory and Experiment, 2012, P02002, doi: 10.1088/1742-5468/2012/00/P00000

Zhang, J., Klingsch, W., Schadschneider, A., & Seyfried, A. 2012, Journal of Statistical Mechanics: Theory and Experiment, 2012, P02002, doi: 10.1088/1742-5468/2012/00/P00000

2012 doi

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

A., Bacik, B

Bacik, K. A., Bacik, B. S., & Rogers, T. 2023, Science, 379, 923, doi: 10.1126/science.add8091

doi:10.1126/science.add8091 2023

[2] [2]

2017, Journal of Statistical Mechanics: Theory and Experiment, 2017, 033404, doi: 10.1088/1742-5468/aa620d

Cao, S., Seyfried, A., Zhang, J., Holl, S., & Song, W. 2017, Journal of Statistical Mechanics: Theory and Experiment, 2017, 033404, doi: 10.1088/1742-5468/aa620d

work page doi:10.1088/1742-5468/aa620d 2017

[3] [3]

2016, Physical Review E, 94, 012312, doi: 10.1103/PhysRevE.94.012312

Cao, S., Zhang, J., Salden, D., et al. 2016, Physical Review E, 94, 012312, doi: 10.1103/PhysRevE.94.012312

work page doi:10.1103/physreve.94.012312 2016

[4] [4]

2023, Crowd

Cordes, J., Chraibi, M., & Tordeux, A. 2023, Crowd

work page 2023

[5] [5]

Dynamics, Volume 4: Analytics and Human Factors in Crowd Modeling, 143, doi: 10.1007/978-3-031-46359-4_6

work page doi:10.1007/978-3-031-46359-4_6

[6] [7]

K., & Wood, D

Elliott, J., Simms, C. K., & Wood, D. P. 2012, Accident Analysis & Prevention, 45, 342, doi: 10.1016/j.aap.2011.07.022

work page doi:10.1016/j.aap.2011.07.022 2012

[7] [8]

2017, PloS one, 12, e0169734, doi: 10.1371/journal.pone.0169734

Prattichizzo, D. 2017, PloS one, 12, e0169734, doi: 10.1371/journal.pone.0169734

doi:10.1371/journal.pone.0169734 2017

[8] [9]

2016, Physical Review E, 94, 032304, doi: 10.1103/PhysRevE.94.032304

Feliciani, C., & Nishinari, K. 2016, Physical Review E, 94, 032304, doi: 10.1103/PhysRevE.94.032304

work page doi:10.1103/physreve.94.032304 2016