Users' traffic on two-sided Internet platforms. Qualitative dynamics

Victoria Rayskin

REVIEW 4 major objections 5 minor 10 references

Users' traffic on two-sided Internet platforms. Qualitative dynamics

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A two-sided platform's long-run traffic is fixed by the local slopes of its attachment functions.

desk verdict A compact, honest synthesis of earlier results: useful qualitative framing for platform traffic, but the central theorem is deferred to the author's own prior paper and the data examples are illustrative, not evidential. read the letter →

arxiv 1908.03059 v2 pith:TSLEKMKB submitted 2019-08-06 physics.soc-ph cs.NAmath.DSmath.NA

Victoria Rayskin This is my paper · ORCID

classification physics.soc-phcs.NAmath.DSmath.NA MSC 37C1037C7591B26

keywords two-sidedplatformsusertrafficdynamicsphaseportraitfixedpointclassificationattachmentfunctionssame-sidenetworkeffectsvectorfieldreconstructionsparseregression

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the long-run behavior of user traffic on two-sided Internet platforms can be read off from the shape of two attachment functions, without solving the full dynamics. It models the fraction of users on each side as a two-dimensional autonomous system in which each side's growth is driven by the opposite side's presence minus its own same-side competition. The central result classifies every fixed point by the product of the slopes of the attachment functions at that point compared with the product of the competition coefficients: a smaller product means stable flow toward the point, a larger product means a saddle, and equality means a center manifold. Under this model there are no repelling fixed points and no cycles, so platforms move monotonically to one of finitely many equilibria. If the theorem is right, platform owners can predict the qualitative fate of the platform from local measurements of user preferences and competition.

What carries the argument

The object that carries the argument is the two-dimensional system $b'=V(g)-\epsilon b$, $g'=W(b)-\delta g$ on the unit square, with $V,W:[0,1]\to[0,1]$ the cross-side attachment functions and $-\epsilon b$, $-\delta g$ the same-side competition terms. The load-bearing identity is the comparison between $V'(g_0)W'(b_0)$ and $\epsilon\delta$ at a fixed point, which decides whether the linearized flow is attracting, repelling-in-one-direction (saddle), or center. A second mechanism is the reconstruction procedure: approximate the vector field by sparse regression over a library of simple functions, using data points from many trajectories, then analyze the recovered equations instead of the raw time series. The classification theorem is what converts the reconstructed vector field into qualitative predictions about basins of attraction and long-run tendencies.

What would settle it

Collect real traffic data from a seller-buyer platform with negative same-side effects and fit the model $b'=V(g)-\epsilon b$, $g'=W(b)-\delta g$. If a fixed point with $V'(g_0)W'(b_0)<\epsilon\delta$ is observed to repel nearby trajectories, or if sustained oscillating user counts appear despite the negative-divergence assumption, the classification would be contradicted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a complete local classification of the phase portrait for the smooth two-sided platform model $b'=V(g)-\epsilon b$, $g'=W(b)-\delta g$: a fixed point $(b_0,g_0)$ is a stable node or spiral exactly when $V'(g_0)W'(b_0)<\epsilon\delta$, a saddle exactly when the product is larger, and lies on a center manifold exactly when the product equals $\epsilon\delta$. Because the divergence is negative, every trajectory in the unit square converges to a fixed point, so the long-run tendency of any 'seller-buyer' type platform is decided by these local slope products. The paper also claims that reconstructing the vector field from observed traffic data, rather than fitting trajectories directly, gives better short-term predictions and reveals the underlying mechanism; this is demonstrated on the GNP/M1 example, where the reconstructed system outperforms the VAR(3) baseline.

Load-bearing premise

The model assumes that every external event, policy change, or shock only moves the platform to a new initial condition, leaving the attachment functions and competition coefficients unchanged.

Editorial extensions

If this is right

Platform owners can identify which equilibria are attracting, which are saddles, and which lie on center manifolds by estimating only the local slopes of the attachment functions at the fixed points.
Since there are no repelling fixed points and no cycles, a stable seller-buyer platform cannot show persistent oscillations; it must settle into one of finitely many steady states.
Basins of attraction are separated by separatrices through saddle points, so small external shocks move traffic to a different steady state only when they push the system across such a separatrix.
For platforms without same-side competition, such as Wikipedia in the paper's model, user volume never decreases on the long run, and observed decline should be attributed to temporary external jumps to lower trajectories.
The reconstruction method applied to real data yields a differential equation whose qualitative analysis can be used for short-term prediction and for planning platform policy.

Reading between the lines

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

Extending the paper's logic, real policy changes and platform redesigns that alter the attachment functions would require re-estimating the vector field; the phase portrait is not invariant under such changes.
If the theorem is combined with data on user preferences, it suggests a testable diagnostic: platforms whose fitted slope product is near $\epsilon\delta$ should exhibit slow or non-generic dynamics, while larger gaps imply faster convergence.
The same classification should transfer to any two-sided market with negative same-side effects and smooth attachment curves, so the result may apply to offline markets such as labor markets and dating markets.
The reconstruction step is where model error enters: if the library of candidate functions omits the true mechanism, the recovered vector field may preserve the phase portrait qualitatively even if it differs term by term, as the paper's own example suggests.

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

Jake — quick take. This is a compact qualitative synthesis, not a new mathematical result. The platform model and Theorem 1 come from the author's earlier papers, and the reconstruction is SINDy with a multi-trajectory twist. That said, the paper is useful: it packages the phase-portrait logic for two-sided platforms clearly, shows basin-of-attraction pictures, and makes the practical point that a platform manager can think in terms of fixed points and separatrices. The reconstruction examples are honest simulations, and the GNP-M1 comparison, while in-sample and without error bars, at least shows the vector-field approach can beat VAR(3) on short-term RMSE in that one case. The writing is plain and the limitations are not hidden: the external-effects-as-initial-conditions assumption is stated up front, and the Wikipedia discussion explicitly hedges on edit wars being temporary.

Soft spots. Theorem 1 is stated without proof and deferred to [9]; that makes the central claim non-self-contained, and the fact that [9] is the author's own paper raises the burden. The stability conditions are plausible from the Jacobian of the planar system, and the divergence argument for no cycles is correct, but a referee cannot verify the center-manifold statement without [9]. The reconstruction sections report no noise levels, regularization parameters, or error bars, so the examples are illustrative rather than evidence. The GNP-M1 result is in-sample; calling it better than VAR(3) is acceptable only as a demonstration, not as a general claim. Finally, the stress-test concern is real: if policy changes or redesigns alter V, W, ε, or δ rather than just the initial condition, the fitted autonomous vector field is a mixture and the phase portrait is not the platform's. The paper acknowledges this only implicitly; a serious revision should state it as a boundary condition.

Who benefits: readers who want an accessible entry to dynamical-systems thinking for platform traffic. It deserves a serious referee because the qualitative claims are testable and the synthesis is coherent, but the referee should demand the proof of Theorem 1 be included or fully specified and the empirical claims relabeled as illustrations.

Referee Report

4 major / 5 minor

Summary. This paper proposes a dynamical-systems framework for studying user traffic on two-sided Internet platforms. It argues that reconstructing the governing vector field from data is preferable to estimating individual trajectories, and it outlines a SINDy-style sparse regression that uses data from many trajectories rather than one. The main theoretical result, Theorem 1 in Section 5.1, classifies the fixed points of the 'seller-buyer' model b'=V(g)-epsilon b, g'=W(b)-delta g according to whether V'(g0)W'(b0) is less than, greater than, or equal to epsilon delta. The paper then applies this framework to power-law and step-function attachments and to Wikipedia-like platforms without same-side network effects, drawing qualitative conclusions about long-run behavior, basins of attraction, and the absence of cycles.

Significance. Theorem 1, if backed by a complete proof, gives a clean and useful criterion: for a plausible two-sided platform model, the long-run behavior is governed by the local slopes of the attachment functions, and cyclic traffic is impossible. The classification is consistent with a direct computation of the Jacobian of (6), whose trace is -(epsilon+delta) and whose determinant is epsilon delta - V'(g0)W'(b0), so the central mathematical claim appears sound. The reconstruction idea of using multiple trajectories in sparse regression is attractive and works in the simulated noiseless examples. However, the paper's applied conclusions for real platforms are not supported by real data, and the main theorem is cited from the author's own previous work rather than derived here. The paper is a useful conceptual contribution, but it needs substantial strengthening in proof detail and in empirical validation before it meets the standards of a journal article.

major comments (4)

[Section 5.1, Theorem 1] The theorem is the paper's central classification result, yet the proof is entirely deferred to [9], an earlier paper by the same author. A direct Jacobian calculation is short (trace is -(epsilon+delta), determinant is epsilon delta - V'(g0)W'(b0)), and the qualitative conclusions of the paper stand or fall on this result. The manuscript should reproduce at least a proof sketch, including the center-manifold condition, or state the theorem as an assumption with a precise reference to the published proof.
[Section 1 and Section 3] The method assumes that all external events act only as switches of initial conditions within one fixed autonomous vector field. If policy changes or platform redesigns alter the attachment functions V, W or the coefficients epsilon, delta, then the data in Section 3 are generated by a family of vector fields, and the single coefficient matrix A fitted by the sparse regression is a mixture that need not correspond to any actual dynamics. The paper should state this as an explicit limitation and, ideally, offer a diagnostic test for whether a single autonomous system is consistent with the data.
[Section 5.2] The Wikipedia discussion depends on the claim that edit wars are 'temporary external effects.' If they are persistent, the decline documented in [6] could reflect a change in the vector field itself, not a jump between trajectories of a fixed field. The conclusion that Wikipedia's long-run popularity increases is therefore conditional on an unverified assumption. The manuscript should make this condition explicit and discuss how one would detect a change in the vector field.
[Section 4] The reconstruction examples provide no quantitative validation: they use noiseless simulated data, do not report regularization parameters, and give no error bars or accuracy metrics. Since reconstruction is the bridge between raw data and the phase-portrait conclusions, the examples should include noisy data from the same system and, ideally, from perturbed systems, to show that the method's output is robust.

minor comments (5)

[Section 2] There are several typos: 'Lorentz' should be 'Lorenz' (two occurrences), 'hight' should be 'height', and 'homogenious monoms' should be 'homogeneous monomials'.
[Section 5.2, Eq. (10)] The displayed first integral is not correct as written. Along solutions of (9), the conserved quantity is ∫ W(b) db - ∫ V(g) dg = const, since d/dt[∫W(b)db - ∫V(g)dg] = W(b)V(g) - V(g)W(b) = 0. The sign and the constant should be corrected.
[Section 5] The paper says it constructs models for Amazon.com, Homes.mil, and Wikipedia, but no real traffic data are used; the figures and phase portraits come from assumed attachment functions. The authors should clarify that these are stylized illustrations rather than empirical reconstructions.
[Section 4, Figures 3-5] The figure captions mention green squares for initial conditions but do not label the axes; adding axis labels would improve readability.
[Section 5.1, Theorem 1] The phrase 'stable nodes/saddles/saddle-nodes, and stable spirals' is imprecise: a saddle-node is a bifurcation phenomenon, not a fixed-point type of a fixed system. The equality case is better described as a non-hyperbolic fixed point lying on a center manifold.

Circularity Check

1 steps flagged · score 4.0 of 10

Theorem 1's proof is deferred to same-author reference [9], making the central formal claim citation-dependent rather than derived in-paper.

self citation load bearing [Section 5.1, Theorem 1; also Section 5 opening sentence]
"The proof of this result can be found in [9]. ... For more detailed analysis and proof of the results stated in this section see [8] and [9]."

Theorem 1 is the paper's central result: it classifies all fixed points of the two-sided platform model (6). The proof is not reproduced here; the derivation chain stops at the same author's prior paper [9]. The introduction likewise credits the 'new dynamical systems approach' to [8] and [9]. Thus the central formal claim rests on the author's own earlier work rather than on an in-paper derivation. The theorem does follow from the stated ODE by a standard Jacobian calculation and is not fitted to data, so this is a partial, authority-based circularity rather than a constructional equivalence; still, the paper's key theorem is not self-contained.

full rationale

Apart from Theorem 1's deferred proof, the paper's qualitative conclusions are not circular. The stability classification in Theorem 1 is the standard linearization of b'=V(g)-epsilon b, g'=W(b)-delta g: the Jacobian has determinant epsilon*delta - V'(g0)W'(b0), so the stated inequalities directly give stable node/spiral versus saddle versus center-manifold; the basin conclusions in the examples follow from the stated phase portraits and monotonicity, not from data fitting. The GNP-M1 comparison in Section 2 is reported without an explicit train/test split, so I do not treat it as a demonstrated fitted-input-called-prediction. The modeling assumption that external effects are switches of initial conditions is a substantive limitation, not a circular step. The main circularity-type concern is the outsourced proof of Theorem 1 to the same-author reference [9], which is load-bearing for the paper's formal claim but does not make the claim equivalent to its inputs.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central model depends on generic attachment functions V,W and coefficients epsilon,delta, which are not fitted in the paper. The examples introduce alpha and a smoothing width, and the Wikipedia argument adds an ad hoc assumption that edit wars are temporary. No new physical or conceptual entities are postulated. The paper's main mathematical proof is imported from the author's prior [9], which is the largest external burden.

free parameters (3)

epsilon, delta (same-side decay rates) = Unspecified; constrained epsilon, delta >= 1
Coefficients in Eq. (6) measure negative same-side network effects; they set the stability threshold in Theorem 1 and are not fitted to data.
alpha (power attachment exponent) = <=1
Exponent in the example attachment functions V(g)=g^alpha and W(b)=b^alpha, Eq. (7).
smoothing half-width in step attachment = Not specified; denoted delta in Eq. (8)
Width of the smooth transition intervals in the three-step attachment; also conflicts with the decay coefficient delta.

assumptions (6)

domain assumption V and W map [0,1] to [0,1], are nonnegative, and vanish at 0
Used throughout Section 5 to keep trajectories in the unit square and to define cross-side attachment; Section 5 states W,V map [0,1] to [0,1] and W(0)=V(0)=0.
domain assumption Same-side network effects can be represented by linear terms -epsilon b and -delta g
Section 5: 'In all examples discussed below, we approximate the same-side network effects with linear functions.' This linearization is load-bearing for the fixed-point classification.
domain assumption External effects are equivalent to changes in initial conditions of a fixed autonomous ODE
Section 1 says externalities shift the process from one trajectory to another, leaving the governing law constant. If false, basin analysis may not apply.
ad hoc to paper Wikipedia's edit wars are temporary external effects, not a persistent same-side term
Section 5.2: 'If the edit wars are not fundamental characteristics... they can be viewed as temporary external effects.' This assumption drives Conclusion 2.
ad hoc to paper Trajectories that escape the unit square do so slowly enough to be ignored over reasonable time scales
Section 5.2: 'we assume that it takes very long time to escape the unit square.' This prevents the model from being rejected when monotone flows leave the domain.
standard math The platform vector field is smooth and has finitely many fixed points
Theorem 1 assumes smoothness and finite fixed points; standard ODE background is used for the stability classification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Users' traffic on two-sided Internet platforms. Qualitative dynamics." pith.science (2026). https://pith.science/paper/TSLEKMKB

@misc{pith2026190803059,
  author       = {Pith},
  title        = {Pith review of: Users' traffic on two-sided Internet platforms. Qualitative dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TSLEKMKB}},
  note         = {Machine review of arXiv:1908.03059}
}

read the original abstract

Internet platforms' traffic defines important characteristics of platforms, such as pricing of services, advertisements, speed of operations. One can estimate the traffic with the traditional time series models like ARIMA, Holt-Winters, functional and kernel regressions. When using these methods, we usually smooth-out noise and various external effects in the data and obtain short-term predictions of processes. However, these models do not necessarily help us to understand the underlying mechanism and the tendencies of the processes. In this article, we discuss the dynamical system approach to the modeling, which is designed to discover the underlying mechanism and the qualitative properties of the system's phase portrait. We show how to reconstruct the governing differential equations from data. The external effects are modeled as system's parameters (initial conditions). Utilizing this new approach, we construct the models for the volume of users, interacting through Internet platforms, such as "Amazon.com", "Homes.mil" or "Wikipedia.org". Then, we perform qualitative analysis of the system's phase portrait and discuss the main characteristics of the platforms.

Figures

Figures reproduced from arXiv: 1908.03059 by the authors.

**Figure 1.** Lorentz attractor Many processes, for which the state variables’ equations cannot be explicitly written due to their complexity, can be approximated with simple differential equations, which have very few non-linear terms. Our goal is to study the complex processes with the help of differential equations that can be fitted into the data, generated by the processes. This approach allows to understand the general l… view at source ↗

**Figure 2.** GNP and Money Supply quarterly data for 1955-1985 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** True and identified trajectories, defined by the Equations (2). All tra [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: True and identified trajectories, defined by the Equations (3). All [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Trajectories of the true Equations (4) and trajectories of the approxi [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: The power attachment functions. The fixed points of the dynamical [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: Smoothed step attachment (with 3 steps) creates 3 attracting fixed [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

**Figure 8.** Figure 8: Fixed points are denoted by red dots. Green separatrix, passing [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

**Figure 9.** Figure 9: If we assume that negative same-side network effect is not present, the [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

**Figure 10.** Figure 10: If the attachment function V (g) has three zeros (at 0, g1 and g2) and the attachment function W(b) has 3 zeros (at 0, b1 and b2), the system has 9 stationary points shown in red color. The flow of users (with some positive volume of users on each side) increases app…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 8 canonical work pages

[9]

Rayskin, Users’ dynamics on digital platforms , Mathematics and Com- puters in Simulation , 142, (2017)

V. Rayskin, Users’ dynamics on digital platforms , Mathematics and Com- puters in Simulation , 142, (2017)

work page 2017
[6]

Halfaker, R.S

A. Halfaker, R.S. Gieger, J. Morgan, J. Riedl The Rise and Decline of an Open Collaboration System: How Wikipedia’s reaction to sudden pop- ularity is causing its decline (2013) American Behavioral Scientist, 57(5) 664-688, DOI:10.1177/0002764212469365

doi:10.1177/0002764212469365 2013
[1]

Beschastnikh, T

I. Beschastnikh, T. Kriplean, D.W. McDonald Wikipedian Self- Governance in Action: Motivating the Policy Lens , ICWSM, 2008

work page 2008
[2]

Brunton, J.L

S.L. Brunton, J.L. Proctorb, J.N. Kutzc, Discovering governing equations from data by sparse identiﬁcation of nonlinear dynamical systems, Pro- ceedings of the National Academy of Sciences, 113, (2016), 3932–3937

work page 2016
[3]

Cowpertwait, P. S. P., & Metcalfe, A. V. (2009) Introductory time series with R. New York: Springer-Verlag 14

work page 2009
[4]

Eisenmann, G

T.R. Eisenmann, G. Parker, M. van Alstyne, Strategies for Two-Sided Markets, Harvard Business Review, 84(10), (2006)

work page 2006
[5]

Halfaker, A

A. Halfaker, A. Kittur, R. Kraut, J. Riedl, it A Jury of Your Peers: Quality, Experience and Ownership in Wikipedia (2009). WikiSym Article 15, 10 pages. DOI:10.1145/1641309.1641332

arXiv 2009
[7]

Krieplean, I

T. Krieplean, I. Beschastnikh, D.W. McDonald, S.A. Golder Community, Consensus, Coercion, Control: CS*W or How Policy Mediates Mass Pop- ulation, presentation at GROUP 2007

work page 2007

Show all 10 references

[8]

Rayskin, Dynamics of two-sided markets, Review of Marketing Science, 14, (2016), 1–19

V. Rayskin, Dynamics of two-sided markets, Review of Marketing Science, 14, (2016), 1–19

2016
[10]

Rochet, J

J.-C. Rochet, J. Tirole, Platform competition in two-sided markets, Jour- nal of the European Economic Association, 1(4), (2003), 990–1029. 15

2003

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [9]

Rayskin, Users’ dynamics on digital platforms , Mathematics and Com- puters in Simulation , 142, (2017)

V. Rayskin, Users’ dynamics on digital platforms , Mathematics and Com- puters in Simulation , 142, (2017)

work page 2017

[2] [6]

Halfaker, R.S

A. Halfaker, R.S. Gieger, J. Morgan, J. Riedl The Rise and Decline of an Open Collaboration System: How Wikipedia’s reaction to sudden pop- ularity is causing its decline (2013) American Behavioral Scientist, 57(5) 664-688, DOI:10.1177/0002764212469365

doi:10.1177/0002764212469365 2013

[3] [1]

Beschastnikh, T

I. Beschastnikh, T. Kriplean, D.W. McDonald Wikipedian Self- Governance in Action: Motivating the Policy Lens , ICWSM, 2008

work page 2008

[4] [2]

Brunton, J.L

S.L. Brunton, J.L. Proctorb, J.N. Kutzc, Discovering governing equations from data by sparse identiﬁcation of nonlinear dynamical systems, Pro- ceedings of the National Academy of Sciences, 113, (2016), 3932–3937

work page 2016

[5] [3]

Cowpertwait, P. S. P., & Metcalfe, A. V. (2009) Introductory time series with R. New York: Springer-Verlag 14

work page 2009

[6] [4]

Eisenmann, G

T.R. Eisenmann, G. Parker, M. van Alstyne, Strategies for Two-Sided Markets, Harvard Business Review, 84(10), (2006)

work page 2006

[7] [5]

Halfaker, A

A. Halfaker, A. Kittur, R. Kraut, J. Riedl, it A Jury of Your Peers: Quality, Experience and Ownership in Wikipedia (2009). WikiSym Article 15, 10 pages. DOI:10.1145/1641309.1641332

arXiv 2009

[8] [7]

Krieplean, I

T. Krieplean, I. Beschastnikh, D.W. McDonald, S.A. Golder Community, Consensus, Coercion, Control: CS*W or How Policy Mediates Mass Pop- ulation, presentation at GROUP 2007

work page 2007