{"id":"fd0aa8c8-5f7c-4adb-81e2-bac10184a45d","arxiv_id":"2505.04032","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"On Erdos-Renyi networks the Extended Yard Sale model shows a continuous transition to local, not global, wealth condensation, with the richest agent's share falling as N^{-1/4} at small nonzero taxation.","lead":"This paper places the Extended Yard Sale wealth-exchange model on random networks and finds that wealth condenses locally instead of globally: the richest agents avoid being neighbors, their fortunes grow, but no single agent takes all. A generalist might read it for a phase-transition view of inequality in which the trading network itself controls the outcome.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 13's N^{-1/4} law is fixed by a visual finite-size collapse with degenerate exponents and no error bars; the 'only fully connected' corollary also overstates what Eq. 13 implies.","rationale":"The reader's weakest assumption correctly identifies the fitted scaling exponents as the least secure support for the central quantitative claim. I agree that Eq. 13, the inference about fully connected networks, and the N^{3/4} rich-wealth scaling in Appendix B all inherit that fit. I add a sharper internal check: the paper itself derives <x1>=N^{-1/4}g_T(kbar/N) with g_T diverging as (1-p)^{-1/4}; this yields <x1>=O(1) for p=1-c/N, i.e., for near-complete ER graphs, so the sentence 'possible only for fully connected networks' is not exactly what Eq. 13 says. This does not overturn the qualitative conclusion for finite-density ER networks, which is supported by direct Monte Carlo data and by the independent result of Burgers and Greengard, nor does it change the need for a conditional verdict. The concrete test with larger N and error bars would settle whether the exponent is really -1/4 and whether the 'only fully connected' corollary should be softened to 'only when 1-kbar/N = O(1/N)'.","tokens_in":15925,"tokens_out":15990,"duration_ms":169891,"concrete_test":"Use QMF (cheap) at T=0.06 with fixed p=kbar/N=0.5 for N=64,128,256,512,1024; average over enough ER samples to produce bootstrap error bars and fit log <x1> vs log N. If the slope is not negative, the central local-condensation claim fails; if it is negative but not close to -1/4, Eq. 13 needs revision. As a secondary check, run the same computation at p=1-2/N: Eq. 13 predicts an O(1) richest-agent share even though the graph is not complete, which would falsify the paper's 'only fully connected' phrasing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative form of the central claim rests on Sec. III.B's finite-size scaling. The exponents are obtained 'by visual inspection' in Fig. 3: nu=mu=infinity at T=0 and nu=mu=-4 at T=0.06, with MC/QMF data only up to N=256 and no error bars. The collapse variable is N^{-nu/mu} kbar, so the data constrain only the ratio nu/mu; the prefactor exponent 1/mu, which fixes the N^{-1/4} decay in Eq. 13, is not independently established. The function g_T(p)=-1/3+p^{1/3}(1-p)^{-1/4} is likewise a phenomenological fit, and Eq. 14 is imposed by requiring the fully connected limit to be recovered, not derived from the networked dynamics. Consequently the statement that only fully connected networks can yield a nonzero <x1> as N->infinity is conditional on these fits. In fact, even granting Eq. 13, p=1-c/N with constant c gives <x1>=O(1), so the paper's 'only fully connected' wording is stronger than its own equation. The qualitative existence of local condensation is independently supported by direct MC and by reference [41]; the concern targets the claimed exponent and its universality across modes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces a networked version of the Extended Yard Sale (EYS) model on Erdős–Rényi random graphs, combining Monte Carlo simulations, a Quenched Mean Field approximation, and a two-population Mean Field theory. The central claim is that, unlike the fully connected case, the networked model does not exhibit global wealth condensation but instead local wealth condensation, in which a set of mutually non-adjacent rich agents holds a submacroscopic share each, while the model still undergoes a continuous phase transition at a temperature Tc. The paper derives the QMF dynamical equations, reduces them to a two-population ODE, and proposes a finite-size scaling law for the expected wealth of the richest agent, x1 ~ N^{-1/4} g_T(kbar/N) at small nonzero temperature, concluding that nonzero macroscopic shares are possible only in fully connected networks. The qualitative claim is supported by simulations and by comparison with the prior result of Börgers and Greengard, but the quantitative scaling rests on exponents and a scaling function obtained by visual inspection and by fitting.","tokens_in":16021,"tokens_out":2871,"duration_ms":29997,"significance":"If the quantitative claims held rigorously, this would be a valuable extension of a well-known econophysics model to structured populations, showing that network topology prevents single-agent oligarchy while preserving a continuous transition. The paper's strengths include a clear formulation of the networked dynamics, a transparent QMF derivation, a two-population MF that recovers the fully connected limit, and a comparative study of four interaction/taxation modes. The qualitative finding of local condensation agrees with an existing rigorous theorem, which lends credibility to the overall picture. However, the quantitative content—especially the N^{-1/4} law and the N^{3/4} rich-agent wealth scaling—is not derived from first principles and is supported only by finite-size collapses with no error bars and with data up to N=256. Those quantitative statements, rather than the mere existence of local condensation, are the main new contributions claimed, so the current support is insufficient for the strength of the abstract and conclusions.","major_comments":[{"comment":"The scaling law in Eq. 13 is the quantitative backbone of the central claim, but the exponents are obtained 'by visual inspection' with no error bars, and only the ratio nu/mu is constrained by the data collapse variable N^{-nu/mu} kbar; the prefactor exponent 1/mu, which fixes the N^{-1/4} decay, is not independently established. With Monte Carlo data only up to N=256, the asymptotic decay is not robust. The authors should either derive this exponent, provide a statistically grounded collapse with error estimates, or explicitly moderate the claim from a determined law to a fitted finite-size observation.","section":"III.B, Fig. 3 and Eq. 13"},{"comment":"The statement that a nonzero average <x1> as N -> infinity 'is possible only for fully connected networks' is stronger than what Eq. 13 implies. Even if Eq. 13 holds, taking p = 1 - c/N with constant c gives kbar/N -> 1 and <x1> = O(1) for any g_T(1) finite, so the equation itself allows macroscopic shares on dense non-fully-connected sequences. The wording should be revised to describe the limit kbar/N -> 1 rather than the exclusive case kbar = N-1.","section":"III.B, paragraph after Eq. 13"},{"comment":"The comparison in Fig. 5c against 'predictions of Eq. 17' is circular: the value R* used to evaluate Eq. 17 is read from the very same MC/QMF simulations whose u* is being fit. The same issue applies to the critical temperature Tc = N/(N-R*). These comparisons demonstrate consistency with the two-population ansatz, but they are not independent tests of the MF theory. The text should label them as consistency checks and avoid the word 'prediction'.","section":"III.B, Fig. 5c and Eq. 17"},{"comment":"The claimed N^{3/4} scaling for the average wealth of the rich agents is presented as a guide-to-the-eye fit without error bars, a derivation, or a scaling collapse analogous to Fig. 3. Since this scaling is presented as a concrete manifestation of local condensation, it needs either an analytical argument or a more careful finite-size analysis with uncertainties; as written, it does not support the conclusion that this exponent is universal across modes and network sizes.","section":"Appendix B, Fig. 10"}],"minor_comments":[{"comment":"The sentence 'Appendix A sheeds light on the origin of this discrepancy' contains a typo: 'sheeds' should be 'sheds'.","section":"III.B, first paragraph"},{"comment":"The phrase 'the heathing phase' should be 'the heating phase'.","section":"Appendix A, Fig. 7 caption"},{"comment":"The phrase 'Without loosing generality' should be 'Without loss of generality'.","section":"II.E, paragraph on taxation modes"},{"comment":"The convention aii in {0,2,4,...} for non-directed networks is unusual and could confuse readers; consider a brief clarification that it encodes the factor 2 for self-loops.","section":"II.B"},{"comment":"The inset of Fig. 5d reports that MC tends to overestimate R* due to stochastic fluctuations; this is an important caveat and should be stated in the main text near the discussion of Fig. 5c as well, since it affects how seriously the consistency between MC and QMF should be taken.","section":"III.B, Fig. 5d"},{"comment":"The derivation of Eq. 6 jumps from the full expectation to the factorized form without explicitly stating the factorization assumption P(wi,wj|a) ≈ P(wi|a)P(wj|a) at the point where the noise average is taken; making that step explicit would improve readability.","section":"II.D, Eq. 6"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the qualitative conclusion is believable, especially given its agreement with the rigorous result of Börgers and Greengard. The main issue is that the quantitative claims—Eq. 13, the 'only fully connected' corollary, and the N^{3/4} scaling—are supported by visual fits without error bars, and the comparison in Fig. 5c is partly circular. I do not see grounds for rejection, but the authors need to either strengthen the statistical analysis or substantially soften the quantitative statements. I would encourage the editor to ask for a revised version that addresses the load-bearing scaling issue head-on."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the paper's qualitative headline—local, not global, condensation on ER networks—is credible and consistent with the Borgers-Greengard preprint [41]. The quantitative wrapping around it, specifically the N^{-1/4} law and the 'only fully connected' corollary, is fitted, not derived, and the corollary is overstated even on the paper's own equations.\n\nThe genuinely new pieces are worth crediting. The networked EYS equation (Eq. 5) is a natural extension, the QMF derivation (Eq. 8) is done carefully, and the two-population mean field (Eqs. 16–17) gives a clean picture of rich–poor coexistence. I checked the fully connected reduction and the two-population algebra as far as the printed equations allow; they are consistent. The paper is also honest: Appendix A openly discusses hysteresis and the unexplained MC/QMF discrepancy in TMB modes, which is more than many papers do.\n\nThe soft spots are real but localized. First, the finite-size scaling exponents (nu = mu = -4 at T = 0.06) are determined by visual inspection with data only up to N = 256 and no error bars. At T = 0 the exponents are reported as infinite, so the collapse fixes only the ratio nu/mu; the prefactor exponent 1/mu that sets the N^{-1/4} decay is not independently pinned down. Second, Eq. 17's 'prediction' uses R* measured from the same MC/QMF data, so it is a self-consistent fit, not a prediction. Third, the claim that only fully connected networks allow O(1) richest-agent wealth is stronger than Eq. 13 actually implies: p = 1 - c/N gives O(1) under that equation. Fourth, the paper never states that reconciling T_c = N/(N-R*) with the observed T_c -> 1 requires R*/N -> 0. None of these kills the qualitative result, which is supported by direct MC and by [41]. But the quantitative package needs disciplined revision. And no code or data is provided, which is a shame for a quantitative scaling claim.\n\nWho this is for: people working on econophysics, wealth condensation, and networked agent-based models. It is a serious contribution, not a throwaway. With revision—proper error analysis, a derived or at least honestly hedged scaling law, and a corrected corollary—it could be a solid paper.\n\nFor peer review: send it out. A referee can fix the scaling claims; the underlying model and the qualitative finding are worth refereeing.","headline":"The qualitative local-condensation result is credible and consistent with prior work, but the N^{-1/4} scaling law and 'only fully connected' corollary are fitted rather than derived, and the corollary overstates the paper's own equation.","tokens_in":16835,"tokens_out":2799,"would_cite":false,"duration_ms":26601,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A networked Extended Yard Sale model leads to local wealth condensation, not a single oligarch, while preserving the continuous phase transition.","keywords":["wealth condensation","Extended Yard Sale model","Erdős–Rényi random networks","random asset exchange","local wealth condensation","phase transition","quenched mean field","econophysics"],"falsifier":"Run the same model at $N=1000$ or larger with many network samples and error bars, and test directly whether the richest agents' wealth grows as $N^{3/4}$ and whether $\\langle x_1\\rangle_a$ collapses under $N^{-1/4}g_T(\\bar{k}/N)$; if the fitted exponents drift with $N$, the local-condensation scaling law is not asymptotic. Deriving the exponents from the Quenched Mean Field equations would also settle it.","tokens_in":15482,"feed_emoji":"💰","tokens_out":9685,"duration_ms":88530,"temperature":0.7,"pith_summary":"The paper introduces a networked version of the Extended Yard Sale model, a random asset-exchange model in which wealth redistribution competes with wealth-attained advantage. On Erdős–Rényi random networks, it argues, the model still undergoes a continuous phase transition, but the condensation is local rather than global: no single agent takes a macroscopic share of total wealth. The load-bearing quantitative claim is a finite-size scaling law: at small nonzero temperature the expected relative wealth of the richest agent scales as $N^{-1/4}g_T(\\bar{k}/N)$, so in the thermodynamic limit a nonzero share for the richest agent is possible only when the network is fully connected. Instead, wealth concentrates in a set of mutually non-adjacent rich agents whose individual wealth grows as $N^{3/4}$, while poor agents hold constant wealth. If this is right, interaction structure—not just the exchange rule—determines whether an economy produces one oligarch or a dispersed rich class.","feed_headline":"Wealth condenses locally, not globally, on random networks","feed_subtitle":"A networked Yard Sale model keeps its phase transition, but spreads the rich into non-adjacent agents.","key_machinery":"The central object is the Quenched Mean Field evolution equation for the expected relative wealth $x_i$ of each node:\n$$\\dot{x}_i = \\frac{T}{N}\\left(-\\beta_i x_i + \\frac{1}{N}\\sum_j \\beta_j x_j\\right) + \\sum_j \\kappa_{ij}(x_i-x_j)\\min\\{x_i,x_j\\},$$\nwith taxation rates $\\beta_i$ and interaction rates $\\kappa_{ij}$ set by the choice of interaction and taxation mode. A two-population reduction splits agents into $R$ rich and $N-R$ poor agents; the marginal-stability condition of that reduced dynamics gives the critical temperature $T_c=N/(N-R^*)$. The quantitative punchline is carried by the scaling hypothesis $\\langle x_1\\rangle_a(\\lambda^\\nu\\bar{k},\\lambda^\\mu N)\\approx \\lambda \\langle x_1\\rangle_a(\\bar{k},N)$, whose data collapse gives $N^{-1/4}\\langle x_1\\rangle_a$ as a function of $\\bar{k}/N$ at small nonzero temperature, together with the observation that rich agents must be non-adjacent, which bounds $R$ through the maximal independent set.","core_discovery":"The central claim is that local wealth condensation replaces global wealth condensation when the Extended Yard Sale model is placed on Erdős–Rényi random networks, while the continuous condensation transition survives. For the four interaction/taxation variants studied, Monte Carlo simulation, Quenched Mean Field approximation, and a two-population Mean Field theory agree that rich agents cannot be neighbors: connected wealthy agents compete until one is ruined, so the rich form an independent set. Quantitatively, for small nonzero temperature $T$, the network-averaged relative wealth of the richest agent obeys $\\langle x_1\\rangle_a \\approx N^{-1/4}g_T(\\bar{k}/N)$, with $g_T(p)=-\\frac{1}{3}+p^{1/3}(1-p)^{-1/4}$ as a phenomenological fit, so a finite fraction of total wealth in one agent's hands survives only as $\\bar{k}\\to N-1$, the fully connected limit. In the condensed phase the average wealth of each rich agent grows as $N^{3/4}$ and that of poor agents stays $O(1)$; the number of rich agents is bounded by the network's maximal independent set, which on Erdős–Rényi graphs gives roughly $2\\ln N/\\ln(1/(1-p))$.","pith_inferences":["If adding links among the rich destroys their advantage, then network rewiring that raises clustering among the wealthy should measurably lower the condensation threshold; this is a testable consequence the paper leaves implicit.","The fitted zero-temperature scaling has degenerate exponents, $\\nu=\\mu=\\infty$ with only $\\nu/\\mu=1$ fixed, so a direct measurement of $\\langle x_1\\rangle_a$ versus $N$ at fixed $\\bar{k}$ would determine whether the $T=0$ law is truly $\\bar{k}/N$ or only an apparent collapse.","The hysteresis loops reported for some taxation modes imply the stationary condensed state is not unique; policy conclusions drawn from uniform-initial-condition runs may miss trajectories in which the richest agent's share stays high.","Extending the same local-competition mechanism to scale-free or small-world topologies, where hubs cannot be adjacent yet dominate connectivity, may yield a different condensation scale than $N^{3/4}$; that is a natural next test."],"forward_implications":["None of the four interaction/taxation modes on Erdős–Rényi networks produces a single agent holding a macroscopic share of total wealth at nonzero temperature; local condensation is generic across modes.","The condensation transition remains continuous, with critical temperature $T_c\\approx 1$ in the large-$N$ limit, so the critical-point phenomenology of the fully connected model carries over to sparse networks.","At any nonzero temperature the relative wealth of the richest agent vanishes as $N^{-1/4}$ times a function of $\\bar{k}/N$, so a nonzero macroscopic share requires $\\bar{k}=N-1$.","Wealth concentrates in an independent set of rich agents: the number of rich agents is bounded by the maximal independent set, roughly $2\\ln N/\\ln(1/(1-p))$ on dense Erdős–Rényi networks.","In the condensed phase individual rich agents hold wealth $\\sim N^{3/4}$ while poor agents hold $O(1)$, producing a two-peaked wealth distribution."],"supporting_citations":[{"why":"Introduces the Extended Yard Sale model with redistribution and wealth-attained advantage; defines the continuous condensation transition this paper generalizes.","marker":"[11]"},{"why":"Introduces the redistribution mechanism that suppresses global condensation, which the networked model inherits.","marker":"[17]"},{"why":"Argues that global wealth condensation is absent in a broad class of networked Yard Sale models and defines local wealth condensation; the paper's main result aligns with and extends it.","marker":"[41]"},{"why":"Provides earlier scaling results connecting local and global wealth condensation on sparse networks that this paper's scaling analysis builds on.","marker":"[40]"},{"why":"Studies Yard-Sale exchange on networks with wealth sharing and appropriation, supplying the prior networked setting.","marker":"[39]"},{"why":"Gives the asymptotic size of the largest independent set in random graphs used to bound the number of rich agents.","marker":"[46]"},{"why":"Supplies the mean-field theory of asset exchange with growth and wealth distribution used to formulate the MF approximation.","marker":"[44]"}],"fun_headline_variants":["Wealth condenses locally on random networks","Yard Sale on networks: rich avoid neighbors, condense locally","Local condensation replaces global in networked Yard Sale","Random graphs change wealth condensation to local scale"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The specific numbers in the central scaling law come from fitting simulation curves, not from a proof, and the largest simulated systems have only 256 agents with no reported error bars.","fun_headline_variants_meta":{"raw":{"variants":["Wealth condenses locally on random networks","Yard Sale on networks: rich avoid neighbors, condense locally","Local condensation replaces global in networked Yard Sale","Random graphs change wealth condensation to local scale"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000564,"raw_usage":{"total_tokens":2722,"prompt_tokens":1039,"completion_tokens":1683,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":1623}},"tokens_in":655,"tokens_out":1683,"duration_ms":12122,"temperature":1.0,"reasoning_tokens":1623,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:44:02.157666+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same model at $N=1000$ or larger with many network samples and error bars, and test directly whether the richest agents' wealth grows as $N^{3/4}$ and whether $\\langle x_1\\rangle_a$ collapses under $N^{-1/4}g_T(\\bar{k}/N)$; if the fitted exponents drift with $N$, the local-condensation scaling law is not asymptotic. Deriving the exponents from the Quenched Mean Field equations would also settle it.","supporting_citations":[{"cited_title":"Chakraborti, I","cited_arxiv_id":null,"evidence_quote":"Introduces the Extended Yard Sale model with redistribution and wealth-attained advantage; defines the continuous condensation transition this paper generalizes."},{"cited_title":"Francisco Cardoso, S","cited_arxiv_id":null,"evidence_quote":"Introduces the redistribution mechanism that suppresses global condensation, which the networked model inherits."},{"cited_title":"Onnela, J","cited_arxiv_id":null,"evidence_quote":"Argues that global wealth condensation is absent in a broad class of networked Yard Sale models and defines local wealth condensation; the paper's main result aligns with and extends it."},{"cited_title":"Newman,Networks(Oxford University Press, Oxford, 2018)","cited_arxiv_id":null,"evidence_quote":"Provides earlier scaling results connecting local and global wealth condensation on sparse networks that this paper's scaling analysis builds on."},{"cited_title":"HereM= P ijaij is preferred since it works for both, directed and non-directed networks","cited_arxiv_id":null,"evidence_quote":"Gives the asymptotic size of the largest independent set in random graphs used to bound the number of rich agents."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the mean-field theory of asset exchange with growth and wealth distribution used to formulate the MF approximation."}],"review_version":1}