{"id":"1005c2d6-4b03-4c06-a521-d8b40ee1dc29","arxiv_id":"2505.03340","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"In a preferential attachment network, a special anomaly node that attracts new edges at fixed rate grows almost linearly in degree, and the degree power-law exponent changes only when the anomaly arrives early.","lead":"A math paper studies what happens to a growing preferential-attachment network when one special node, the anomaly, starts attracting new links with fixed probability. It finds that the anomaly's degree grows almost linearly in network size, and that an early anomaly changes the degree distribution's power-law exponent, while a late one barely matters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The early-anomaly tail in Eq. (18) is derived from post-τ vertices whose maximum expected degree is only t^{b(1−γ)}, so the observed tail in Figure 5 is dominated by the pre-τ cohort, not by the population used to derive Eq. (18).","rationale":"The exact expectation results (Proposition 1 and Section 3) are carefully derived and appear sound. The load-bearing weakness is the passage from expected degrees to the degree distribution in Section 4, specifically the early-anomaly case. The reader identified the unproved mean-degree-to-distribution identification; I sharpen that concern: the heuristic in Section 4.3 discards the pre-τ cohort as a vanishing fraction, but Eq. (10) implies that this vanishing cohort contains all high-degree ordinary vertices. For the parameters of Figure 5, post-τ ordinary vertices have expected degree at most about 2.7, so the plotted tail is not the population from which Eq. (18) was derived. Thus the claimed agreement between Eq. (18) and simulations is not meaningful. Section 5's explanation of the outliers is also based on an incorrect expansion of the oldest-vertex expected degree. This does not refute the fixed-k asymptotic claim, but it means the central early-anomaly exponent is unsupported by the presented numerical evidence and the heuristic's domain is mis-specified. Therefore the paper should remain conditional: the exact parts stand, while the headline distributional claim needs a rigorous derivation or simulations that isolate the post-τ population, include error bars, and address the crossover scale.","tokens_in":10763,"tokens_out":24483,"duration_ms":235559,"concrete_test":"Run an early-anomaly simulation with a sizeable pre-τ cohort, e.g. m=1, δ=0, β=5, t=10^6, τ=10^{5.4} (γ=0.9); exclude the anomaly, keep all ordinary vertices, and estimate the CCDF slope over k∈[3, 100]. Post-τ expected degrees are at most (t/τ)^{1/7}≈1.25, so any power-law tail in this range comes from pre-τ vertices. Eq. (18) predicts a CCDF slope of 3+β/m+δ/m−1=7; the pre-τ prediction is 3+δ/m−1=2. If the fitted slope is near 2, the early-anomaly exponent in Eq. (18) is not the observable tail; if it approaches 7 after conditioning on birth time i>τ and restricting k below the crossover, the heuristic can be salvaged.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 derives Eq. (18), p_k ~ k^{-(3+β/m+δ/m)}, by applying the heuristic only to vertices born after τ=t^γ, because the pre-τ fraction vanishes. But pre-τ vertices cannot be neglected in the tail. From Eq. (10), for i<τ, E[D_i(t)+δ] ≈ (m+δ)(t/τ)^{m/(2m+β+δ)}(τ/i)^{m/(2m+δ)}; from Eq. (16), post-τ expected degrees are (m+δ)(t/i)^{m/(2m+β+δ)}. Thus the largest post-τ expected degree is ~t^{b(1−γ)} with b=m/(2m+β+δ), while pre-τ vertices reach ~t^{b+γ(a−b)}, a=m/(2m+δ)>b. Since b+γ(a−b)>b(1−γ), every ordinary vertex of sufficiently large degree is pre-τ and follows the standard-PA tail 3+δ/m, not Eq. (18). For Figure 5's parameters (t=50000, m=1, δ=0, β=5, γ≈0.3615), post-τ expected degrees are at most ~2.7, so the plotted tail is entirely pre-τ plus the anomaly; Eq. (18) is compared against a different population. Section 5's explanation compounds this by misstating the oldest-vertex exponent: it writes (t/τ)^{1/2+β/m+δ/m}τ^{1/2+δ/m}, whereas Eq. (10) gives (t/τ)^{m/(2m+β+δ)}τ^{m/(2m+δ)}. The fixed-k limit may still be Eq. (18), but the supporting comparison in the early case does not test it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a preferential attachment model in which an anomalous vertex v_tau arrives at time tau and, after arrival, attracts each new edge with an additional fixed probability beta/(2m+beta+delta). The model interpolates between ordinary PA and the superstar model. The authors derive exact recursions for the expected degree of the anomaly and of ordinary vertices (Proposition 1 and Eqs. (9)-(11)), prove almost-sure convergence of suitably rescaled degrees via martingale arguments (Section 3.3), and then give a heuristic derivation of the ordinary-vertex degree-distribution exponent in three regimes: late (tau = t - t^gamma), mid-way (tau = alpha t), and early (tau = t^gamma) arrival. The claimed exponents are 3+delta/m for late and mid-way anomalies and 3+(beta+delta)/m for an early anomaly. The heuristic formulas are compared with simulations in Figures 3-5.","tokens_in":11125,"tokens_out":12876,"duration_ms":122467,"significance":"The exact expectation formulas and the martingale convergence statements are clean and useful, and the model is a natural extension of the superstar model to a general arrival time. The almost-sure convergence results in Section 3.3 are a genuine contribution, as is the closed-form formula for the anomaly's mean degree. If the early-anomaly exponent claim were properly supported, the paper would be an interesting contribution to the study of change-point and anomaly effects in PA networks. However, the numerical support for the early-anomaly case does not test the claimed exponent, because the simulated tail is dominated by vertices born before the anomaly. Thus the central new phenomenon advertised in the abstract remains only a heuristic without valid finite-sample evidence.","major_comments":[{"comment":"The derivation of Eq. (18) applies only to vertices born after tau = t^gamma, since the pre-tau fraction is vanishing. In the simulation of Figure 5 (t=50000, m=1, delta=0, beta=5, gamma approx 0.3615), the largest expected degree of a post-tau vertex is t^{m(1-gamma)/(2m+beta+delta)} approx 2.7. Consequently, every ordinary vertex of degree larger than about 3 was born before tau and follows the expected degree formula in Eq. (10), which produces the standard-PA tail k^{-(3+delta/m)} rather than Eq. (18). The tail shown in Figure 5 is therefore dominated by the pre-tau cohort, and the figure does not test the claimed early-anomaly exponent. The fixed-k limit may well be Eq. (18), but the paper needs either a simulation that isolates post-tau vertices, a parameter regime in which the post-tau expected degrees span a sufficiently large range, or a more explicit discussion of why the finite-t tail cannot be used to read off the fixed-k exponent.","section":"Section 4.3, Eq. (18), Figure 5"},{"comment":"The explanation of Figure 5 attributes the right-deviation of the empirical tail to the oldest vertex's faster degree growth. This is not accurate. From Eq. (10), for every pre-tau vertex i with i < tau = t^gamma, the expected degree satisfies E[D_i(t)+delta] approx (m+delta)(t/tau)^{m/(2m+beta+delta)}(tau/i)^{m/(2m+delta)}, so the entire pre-tau cohort has expected degree larger than the maximum post-tau expected degree, which is only about t^{m(1-gamma)/(2m+beta+delta)}. Thus the outliers to the right in Figure 5 are not exclusively, or even mainly, the oldest vertex; they are the whole pre-tau population. The text in Section 5 should be corrected to state that the finite-t tail is generated by pre-tau vertices and therefore is not a valid comparison for Eq. (18).","section":"Section 5"}],"minor_comments":[{"comment":"The phrase 'converges almost surely as t to infinity and tau to infinity' is imprecise, since tau is fixed as a model parameter. It should read 'for each fixed tau, as t to infinity', or the authors should specify that they consider a sequence with tau tending to infinity.","section":"Section 3.3"},{"comment":"For t=50000, the value tau=50 is only approximately t^gamma with gamma=0.3615; the exact value of 50000^{0.3615} is about 50.1. The caption should state whether tau is taken as floor(t^gamma) or as exactly 50.","section":"Section 4.4, Figure 5"},{"comment":"The Stirling approximation in Eq. (8) is stated for fixed a, but it is later applied to exponents that depend on model parameters such as m/(2m+beta+delta). This is standard and harmless, but a brief note would improve rigor.","section":"Eq. (8)"},{"comment":"References [5] and [8] are given as arXiv preprints; if published versions exist, they should be cited instead of or in addition to the arXiv identifiers.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The exact parts of the paper are solid and the model is worth publishing after revision. The main issue is the early-anomaly section: the heuristic exponent may be correct as a fixed-k limit, but the current simulation comparison does not test it, and the text misinterprets the source of the discrepancy. I would encourage the authors to either add a valid numerical check (e.g., conditioning on post-tau vertices and choosing parameters where their expected degree range is large) or explicitly present Eq. (18) as an unverified conjecture with the finite-t tail explained by pre-tau vertices."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the exact mean-degree analysis is solid and worth having, but the advertised early-anomaly degree-distribution claim is a heuristic that the simulations don't actually test, and Section 5 contains a clear formula error.\n\nThe model extends the superstar by letting the anomaly arrive at arbitrary time, and the recursions in Prop. 1 and Eqs. (10)-(11) are correct and clean. The martingale convergence for ordinary vertices is also fine. The late- and mid-way-anomaly heuristics give the standard PA exponent with a modified amplitude, and the mid-way factor is well supported by the simulations. That is a legitimate incremental contribution.\n\nThe early-anomaly case is the soft spot. Section 4.3 derives p_k ~ k^{-(3+(β+δ)/m)} from the post-τ vertices only. But for the simulated parameters (t=50000, τ=50, m=1, β=5), the largest post-τ expected degree is only about t^{b(1−γ)} ≈ 2.7. The tail in Figure 5 is therefore generated by the pre-τ vertices plus the anomaly, not by the population used in the derivation. The comparison does not test Eq. (18). The authors say the heuristic is non-rigorous, but the figure gives no support either.\n\nSection 5's explanation of the rightward outliers compounds the problem: it states E[D_1(t)+δ] ≈ (t/τ)^{1/2+β/m+δ/m} τ^{1/2+δ/m}, while Eq. (10) gives (t/τ)^{m/(2m+β+δ)} τ^{m/(2m+δ)}. Those differ a lot. For the Figure 5 parameters the correct exponent is about 0.27; the text's version is far larger. This needs a straightforward correction.\n\nNone of this undermines the exact results, and the fixed-k limit in (18) may still be correct. But the paper as posted does not make the case for the early-anomaly exponent, and the simulations have no error bars or code. A referee should ask for simulations at much larger t (or a different scaling) where the post-τ population actually produces the predicted tail, and for the Section 5 formula to be fixed.\n\nThis paper is for the PA/random-graphs community and people thinking about anomaly detection in growing networks. It deserves serious peer review, not a desk reject. I'd send it out expecting the early-anomaly section to be revised.","headline":"Useful exact mean-degree results for a PA model with an arbitrary-time anomaly, but the early-anomaly exponent is heuristic and untested by the simulations, and Section 5 has a formula error.","tokens_in":11681,"tokens_out":9034,"would_cite":true,"duration_ms":75967,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C80","60C05"],"pacs":["89.75.Hc"],"model":"deepseek-v4-flash","headline":"An anomaly that enters a preferential-attachment network early changes the degree distribution's power-law exponent; the same anomaly arriving late leaves almost no trace.","keywords":["preferential attachment","anomalous vertex","degree distribution","power-law exponent","superstar model","expected degree","dynamic network","change point"],"falsifier":"Simulate the model with $m=1$, $\\delta=0$, $\\beta=2$, anomaly at $\\tau=1000$, and run to $t=10^5$ and $t=2\\times 10^5$. Proposition 1 predicts $\\mathbb{E}[D_\\tau(t)+\\delta]/t \\to \\frac{m\\beta}{m+\\beta+\\delta} = \\frac{2}{3}$. If the empirical ratio converges to a different constant, or the finite-time gamma-ratio correction is not observed, the closed form is wrong. For the heuristic distribution, simulate $\\tau=\\alpha t$ with $\\alpha=1/2$ over many runs and compare the tail of pre-anomaly vertices to the predicted slope $-(2+\\frac{\\delta}{m})$ with pre-factor $\\alpha^{\\frac{\\beta}{2m+\\beta+\\delta}}$.","tokens_in":10533,"feed_emoji":"📈","tokens_out":8379,"duration_ms":73649,"temperature":0.7,"pith_summary":"This paper studies a preferential attachment network into which a single 'anomaly' vertex arrives at time τ and thereafter captures a fixed share of every new edge. It proves an exact formula for the anomaly's expected degree, which grows almost linearly with network size, and derives the power-law exponent of the ordinary vertices' degree distribution as a function of τ. The main finding is a timing dichotomy: an early anomaly significantly steepens the degree distribution, while a late anomaly leaves the network nearly indistinguishable from the standard preferential attachment model. The paper also shows that the oldest vertex grows faster under an early anomaly, which explains the visible outliers in the simulated tail.","feed_headline":"Early anomaly steepens a network's degree tail; late one is invisible","feed_subtitle":"If the anomaly arrives early it rewrites the power-law exponent; late arrival barely matters","key_machinery":"The load-bearing object is the recursion for expected degrees under the modified attachment rule, in which the anomaly's edge-receiving probability is $((t-1)\\beta + D_\\tau + \\delta)$ divided by the universal denominator. Solving this recursion yields the exact closed form in Proposition 1 for the anomaly's mean degree. The heuristic degree-distribution argument then uses the asymptotic formula for $\\mathbb{E}[D_i(t)+\\delta]$ as a function of the vertex index $i$ to count how many indices have expected degree near $k$; inverting this index-to-degree map produces the power-law exponents in (14), (17), and (18).","core_discovery":"The central claim is that the presence of an anomalous vertex that attracts a fixed probability of each new edge—on top of its normal preferential-attachment share—changes the network's degree structure in a way that depends sharply on when the anomaly appears. For the anomaly itself, the expected degree obeys the exact identity $\\mathbb{E}[D_\\tau(t)+\\delta] = \\frac{m\\beta t}{m+\\beta+\\delta} + c_0 \\frac{\\Gamma(t + \\frac{m}{2m+\\beta+\\delta})\\Gamma(\\tau)}{\\Gamma(t)\\Gamma(\\tau + \\frac{m}{2m+\\beta+\\delta})}$, so it grows linearly with slope $\\frac{m\\beta}{m+\\beta+\\delta}$, larger than the per-edge attraction probability $\\frac{\\beta}{2m+\\beta+\\delta}$. For ordinary vertices, the paper argues heuristically that the degree distribution is still a power law, with exponent $3+\\frac{\\delta}{m}$ when the anomaly arrives late ($\\tau=t-t^\\gamma$) or mid-way ($\\tau=\\alpha t$), and exponent $3+\\frac{\\beta+\\delta}{m}$ when it arrives early ($\\tau=t^\\gamma$); the mid-way case also picks up a multiplicative factor $\\alpha^{\\frac{\\beta}{2m+\\beta+\\delta}}$.","pith_inferences":["The exact gamma-ratio correction in the anomaly's expected degree quantifies how long it takes the anomaly to reach its linear asymptote; this transient could serve as a finite-time signature of the anomaly's age.","The same index-to-degree heuristic, applied separately to pre- and post-anomaly vertices, would yield the full two-population degree distribution; the paper only states the aggregate.","Comparing the early-anomaly exponent $3+\\frac{\\beta+\\delta}{m}$ with the superstar model's $3+\\frac{p}{1-p}$ could in principle distinguish a constant edge-attraction mechanism from a degree-proportional one in empirical networks."],"forward_implications":["If the anomaly arrives early, the ordinary vertices' power-law exponent jumps from $3+\\frac{\\delta}{m}$ to $3+\\frac{\\beta+\\delta}{m}$, so the network tail becomes considerably thinner.","If the anomaly arrives mid-way, the exponent stays $3+\\frac{\\delta}{m}$ but the degree distribution is multiplied by $\\alpha^{\\frac{\\beta}{2m+\\beta+\\delta}}$, meaning the high-degree pre-anomaly vertices grow slower than in the standard model.","If the anomaly arrives late, the degree distribution converges to the standard preferential attachment power law, so detecting the anomaly from the degree sequence alone becomes hard.","The anomaly's own degree grows linearly at a rate larger than its fixed edge-capture probability, because it also receives edges through the regular preferential attachment channel.","The oldest vertex's expected degree grows as $t^{\\frac{\\gamma}{2+\\delta/m} + \\frac{1-\\gamma}{2+(\\beta+\\delta)/m}}$ for an early anomaly, outstripping the rate implied by the early-anomaly power law and producing the observed right tail."],"supporting_citations":[{"why":"supplies the standard PA degree recursion and martingale convergence toolkit that the exact formulas and the heuristic build on.","marker":"[9]"},{"why":"defines the superstar model, the baseline case where the anomaly is the initial vertex, and gives the modified power-law exponent.","marker":"[7]"},{"why":"provides the 'expected degree in a unit interval implies degree k' heuristic used to derive the degree distribution.","marker":"[11]"},{"why":"defines the preferential attachment rule with fitness parameter $\\delta$ that the model's attachment probabilities use.","marker":"[4]"},{"why":"introduces the preferential attachment model whose power-law degree distribution is the baseline for comparison.","marker":"[2]"}],"fun_headline_variants":["Anomaly's timing dictates network degree distribution shift","Early anomaly rewrites network's power-law exponent","Anomaly degree grows linearly, alters network structure","Late anomaly leaves degree tail intact, early one reshapes it","Preferential attachment anomaly: arrival time is key"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The heuristic derivation assumes that a vertex's degree is tightly concentrated around its expected value, so that the fraction of vertices with expected degree near $k$ can be equated with the true fraction of degree-$k$ vertices; the paper does not prove concentration, and the rigorous derivation is left open.","fun_headline_variants_meta":{"raw":{"variants":["Anomaly's timing dictates network degree distribution shift","Early anomaly rewrites network's power-law exponent","Anomaly degree grows linearly, alters network structure","Late anomaly leaves degree tail intact, early one reshapes it","Preferential attachment anomaly: arrival time is key"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1219,"prompt_tokens":929,"completion_tokens":290,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":214}},"tokens_in":545,"tokens_out":290,"duration_ms":3378,"temperature":1.0,"reasoning_tokens":214,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:54:53.124191+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the model with $m=1$, $\\delta=0$, $\\beta=2$, anomaly at $\\tau=1000$, and run to $t=10^5$ and $t=2\\times 10^5$. Proposition 1 predicts $\\mathbb{E}[D_\\tau(t)+\\delta]/t \\to \\frac{m\\beta}{m+\\beta+\\delta} = \\frac{2}{3}$. If the empirical ratio converges to a different constant, or the finite-time gamma-ratio correction is not observed, the closed form is wrong. For the heuristic distribution, simulate $\\tau=\\alpha t$ with $\\alpha=1/2$ over many runs and compare the tail of pre-anomaly vertices to the predicted slope $-(2+\\frac{\\delta}{m})$ with pre-factor $\\alpha^{\\frac{\\beta}{2m+\\beta+\\delta}}$.","supporting_citations":[{"cited_title":"Volume 1, Cambridge Series in Statistical and Probabilistic Mathematics, vol","cited_arxiv_id":null,"evidence_quote":"supplies the standard PA degree recursion and martingale convergence toolkit that the exact formulas and the heuristic build on."},{"cited_title":"The Annals of Applied Probability25(5), 2462–2502 (2015)","cited_arxiv_id":null,"evidence_quote":"defines the superstar model, the baseline case where the anomaly is the initial vertex, and gives the modified power-law exponent."},{"cited_title":"Nieuw archief voor wiskunde5(24), 103–113 (2023)","cited_arxiv_id":null,"evidence_quote":"provides the 'expected degree in a unit interval implies degree k' heuristic used to derive the degree distribution."},{"cited_title":"In: SODA ’05: Proceedings of the sixteenth annual ACM-SIAM symposium on 16 Q","cited_arxiv_id":null,"evidence_quote":"defines the preferential attachment rule with fitness parameter $\\delta$ that the model's attachment probabilities use."},{"cited_title":"Science 286(5439), 509–512 (1999)","cited_arxiv_id":null,"evidence_quote":"introduces the preferential attachment model whose power-law degree distribution is the baseline for comparison."}],"review_version":1}