{"id":"405e7638-4c17-4307-964e-37cc216f1995","arxiv_id":"1908.01999","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"In a simple preferential attachment network model, replacing the time derivative with a fractional-order derivative makes node activity peak and then decline, which the authors interpret as aging due to memory.","lead":"A physics team used fractional calculus to add memory to a model of how connections form in networks, finding that remembering past links makes nodes slow down and age. This could help explain why old nodes in social or communication networks stop attracting new connections, though the mathematics in the preprint has serious errors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fractional-calculus derivation is internally inconsistent: Eq. (7) is not the Caputo derivative and the Eq. (4) kernel is non-integrable, so the aging claim lacks its stated basis.","rationale":"The paper's goal is to show that incorporating long memory via a Caputo fractional derivative changes the mean-field activity equation so that node activity peaks and declines. For that claim to hold, the passage from Eq. (4) to Eq. (6) must be a valid fractional-calculus step. It is not: the kernel in Eq. (4) is singular of order α−2, so for every 0≤α<1 the convolution integral diverges; and the operator in Eq. (7) is the Riemann–Liouville fractional integral I^{1−α}y, not the Caputo derivative, which requires y' in the integrand. The numerical formula Eq. (9), however, is the standard predictor–corrector discretization of a genuine Caputo initial-value problem, so a corrected derivation might still produce the aging effect. This is why the concern is about the demonstrated basis of the claim rather than about whether the phenomenon can exist. The reader's weakest assumption points to the same equations; my independent check of Eq. (7) with y=1 confirms the problem. Since the manuscript as written contains these load-bearing inconsistencies and no code or data, I do not see a basis to overturn the REJECT verdict; a revised version that states a correct Caputo model and verifies Fig. 3 with it would be a different question.","tokens_in":7833,"tokens_out":13025,"duration_ms":147133,"concrete_test":"Independently implement the predictor–corrector scheme (Eq. 9) for the initial-value problem ^C D^α k = m/N + m k/K using the standard Caputo definition ^C D^α y(t)=1/Γ(1−α)∫_0^t (t−s)^{-α} y'(s) ds, with the same α∈{0.2,0.4,0.6,0.8}, N, m, and t0 as in Fig. 3, and compare k̄_i(t) to Fig. 3(a); if the activity still peaks and declines, the aging claim survives the corrected definition, whereas if it does not, the claim is an artifact of the misdefined Eq. (7).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that memory causes node-activity aging (Fig. 3 caption) rests on deriving Eq. (6) from the memory-kernel equation Eq. (4) and then solving it numerically via Eq. (9). The derivation is not valid as written. First, with κ(t−t')=(t−t')^{α−2}/Γ(α−1) and 0≤α<1, the integral in Eq. (4) is not a fractional integral of positive order: ∫_0^T u^{α−2}du diverges at u=0 for every α<1, so the right side of Eq. (5) is undefined for continuous bracket functions. The operator labelled D^{-(α−1)} is in fact a Riemann–Liouville derivative of order 1−α, not an integral. Second, Eq. (7) defines c D^α_t y as 1/Γ(1−α)∫(t−s)^{−α} y(s) ds; the standard Caputo derivative for 0<α<1 is 1/Γ(1−α)∫(t−s)^{−α} y'(s) ds. Applying Eq. (7) to y(s)=1 gives a nonzero function, whereas the Caputo derivative of a constant is zero. Thus the identity used to convert Eq. (4) into Eq. (6) is not the Caputo calculus the paper invokes. Since the aging prediction is justified as a consequence of this conversion, the stated theoretical basis for the central claim is missing. Note, however, that the predictor–corrector formula Eq. (9) is the standard discretization of the well-posed Caputo problem ^C D^α k=f; independent implementation of that problem could still exhibit the claimed peak-and-decline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a temporal-network model with preferential attachment and memory. In the memoryless case it derives an analytical solution for node activity, identifies a crossover time, and validates the result by simulation. Memory is introduced by replacing the integer-order derivative with an integral equation containing a power-law kernel, which the authors then convert into a Caputo-type fractional differential equation of order α. The fractional equation is solved numerically with a predictor-corrector scheme, and the paper reports that the effective node activity reaches a peak and then declines, interpreting this as an aging effect. It also reports changes in the degree and link-weight distributions for different fractional orders. The central claim is that memory causes decay of node activity and reduces the chance of older nodes to receive new connections.","tokens_in":8225,"tokens_out":5714,"duration_ms":106007,"significance":"If the derivation were mathematically sound, the paper would offer a compact fractional-calculus framework for introducing long-term memory into temporal-network models and a concrete, testable prediction of aging. Credit is due for the correct memoryless analytical solution, the identified crossover time, the straightforward simulation of the ordinary differential model, and the use of a standard predictor-corrector algorithm. However, the central derivation from the memory-kernel equation to the fractional differential equation contains fundamental mathematical errors. As written, the aging claim is not established as a consequence of the proposed memory mechanism, and the numerical experiments solve a fractional initial-value problem whose connection to the original network model is not demonstrated.","major_comments":[{"comment":"Equation (7) defines the Caputo derivative of order α as 1/Γ(1−α) times the integral of (t−s)^{−α} y(s) ds. This is not the standard Caputo derivative for 0<α<1, which instead has y′(s) in the integrand. With the definition as written, the fractional derivative of a constant is nonzero (it equals (t−t0)^{1−α}/Γ(2−α)), contradicting the property of Caputo derivatives on which the paper relies. Because Eq. (6) is obtained by applying this operator, the conversion from Eq. (4) to Eq. (6) is invalid.","section":"Sec. 2, Eq. (7)"},{"comment":"The memory kernel κ(t−t′)=(t−t′)^{α−2}/Γ(α−1) is non-integrable for all 0≤α<1: the exponent α−2 is less than −1, so the integral of u^{α−2} diverges at u=0. Consequently the integral in Eq. (4) is not a well-defined fractional integral of positive order for continuous integrands, and the notation cD^{−(α−1)} in Eq. (5) is not justified. The paper's premise that this kernel yields a fractional integral equation is load-bearing and false as stated.","section":"Sec. 2, Eq. (4)"},{"comment":"Even setting aside the kernel singularity, the operator manipulation from Eq. (5) to Eq. (6) is inconsistent. Equation (5) contains D^{−(α−1)} acting on the bracket after dki/dt, while Eq. (6) states D^α ki equals the bracket. The required identities from fractional calculus (such as the composition of fractional integral and derivative operators) are not provided, and with the definitions used in the paper they do not hold. Thus Eq. (6), the equation actually solved numerically, is not derived from the memory model in Eq. (4).","section":"Sec. 2, Eqs. (5)–(6)"},{"comment":"The numerical scheme in Eq. (9) is the standard predictor-corrector discretization of a Caputo-type initial-value problem, and the peak-and-decline in Fig. 3 may be a genuine property of that fractional equation. However, the paper does not establish that this initial-value problem is equivalent to the memory mechanism defined by Eq. (4). The abstract and Fig. 3 caption attribute the decline to 'the aging process' and memory, but the theoretical basis for that attribution is missing. This is the central claim of the paper, so the error is not a minor presentation issue.","section":"Sec. 2, Fig. 3 and Eq. (9)"}],"minor_comments":[{"comment":"There is a typo: 'momory' should be 'memory'.","section":"Sec. 2, paragraph before Eq. (4)"},{"comment":"The integral notation in Eq. (4) is confusing: the differential dt′ appears after the kernel rather than multiplying the bracket; the intended expression is κ(t−t′)[m/N + mki(t′)/∑kj(t′)]dt′.","section":"Sec. 2, Eq. (4)"},{"comment":"The text says 'applying a fractional Caputo derivative of order α−1' but then writes Eq. (6) with order α; the stated and used orders do not match.","section":"Sec. 2, Eqs. (5)–(6)"},{"comment":"The characteristic time is written as t∗=c²N²/(4m²) in the text but as t∗=c²N²/(4m) in the Fig. 2 caption; one of these is inconsistent.","section":"Sec. 2, Fig. 2 and text after Eq. (3)"},{"comment":"The sentence 'The BA model B predicts that after a transient period the connectivity distribution of all nodes becomes a Gaussian around its mean value' is incorrect: the Barabási–Albert model produces a power-law degree distribution, not a Gaussian. This affects the interpretation of the deviation reported in Fig. 5(a).","section":"Sec. 2, Fig. 5 discussion"},{"comment":"The caption says 'with as initial condition m=10 nodes and every new node connecting to earlier nodes,' which is inconsistent with the fixed-N model described in Sec. 1 and with the simulation parameters N=1000 reported elsewhere.","section":"Sec. 2, Fig. 3 caption"}],"recommendation":"reject","confidential_remarks":"The reader's stress-test concern is correct and lands directly on the central claim. The fractional-calculus derivation contains definitional and operator-manipulation errors that cannot be repaired by local edits: the memory kernel in Eq. (4) is non-integrable for the stated range of α, Eq. (7) is not the Caputo derivative, and Eq. (6) is not a valid consequence of the memory model. The numerical results might be salvageable if the authors instead propose a fractional differential equation as a phenomenological model, but that is not what the paper claims to do. I recommend rejection; a substantially rewritten manuscript with a correct fractional formulation could be considered anew."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the memoryless preferential-attachment model in Eqs. (1)-(3) is solved correctly, and the simulation matches the analytic curve in Fig. 2. That part is genuinely fine. Second, the fractional-calculus section that is supposed to add memory is mathematically flawed in ways that invalidate the central claim that memory causes node-activity aging.\n\nWhat is actually new: the idea of coupling node-activity dynamics to a power-law memory kernel via fractional calculus is a reasonable thing to try, and the paper does make the memoryless limit alpha=1 reduce to the BA-like behavior. The qualitative observation that stronger memory (smaller alpha) leads to a peak-and-decline in effective node activity is interesting, if it were supported. But the derivation does not support it.\n\nThe soft spots are real and load-bearing. In Eq. (4), the kernel κ(t-t′) = (t-t′)^{α-2}/Γ(α-1) with 0 ≤ α < 1 is not integrable: the integral diverges at t′=t for every α in that range. So calling the right side a fractional integral of order α−1 is wrong; the order is negative, and the expression is undefined as written. Equation (7) for the Caputo derivative omits the derivative of y; the standard Caputo definition for 0<α<1 has y′(s) in the integrand, not y(s). Therefore the operator used to go from Eq. (4) to Eq. (6) is not the Caputo derivative, and the identity claimed does not hold. These are not typos. The entire aging prediction is justified by converting the memory-kernel equation into a Caputo differential equation, and that conversion is invalid.\n\nOne nuance: the predictor-corrector scheme in Eq. (9) is the standard discretization of a well-posed Caputo problem of the form ^C D^α k = f. If one simply posits that Caputo equation as an ad hoc model, the numerical solution might indeed show the reported peak-and-decline. But the paper does not derive that equation from the memory-kernel model, and the simulations are presented as validating a derivation that does not hold. There is also no empirical data, no code, and no parameter fitting; the degree-distribution analysis is qualitative. Citation to prior work, including the same group's Safdari et al., makes clear that the aging phenomenon is not new even if this specific fractional formulation is.\n\nThis paper is not ready for serious peer review as written. It has a nice memoryless starting point and a plausible qualitative idea, but the core mathematical step is broken and no independent evidence is offered. My recommendation: reject. If the authors reframe the model as an ad hoc fractional differential equation, validate it against real temporal-network data, and provide code, it could be resubmitted as a different paper.","headline":"The memoryless part is clean and correct, but the fractional-calculus derivation is broken, so the aging claim rests on air.","tokens_in":8738,"tokens_out":2871,"would_cite":false,"duration_ms":30268,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Memory in temporal networks makes node activity peak and then decline with age.","keywords":["memory effects","temporal networks","link formation","fractional calculus","Caputo derivative","preferential attachment","node activity","aging"],"falsifier":"Check whether the kernel (t−t')^{α−2}/Γ(α−1) is integrable on [t0,t] for α<1, or solve the same preferential-attachment equation using the standard Caputo derivative (with the first derivative of node activity inside the integrand) for α=0.5 and compare the resulting activity curve to the paper's Fig. 3.","tokens_in":7623,"feed_emoji":"🕸️","tokens_out":8136,"duration_ms":86018,"temperature":0.7,"pith_summary":"This paper tries to establish that long-term memory in temporal networks changes who receives new links: if nodes remember past interactions and favor them, a node's activity no longer grows without bound but rises to a peak and then declines. The authors build a preferential attachment model in which node activity follows an ordinary rate equation, then replace the instantaneous derivative with a fractional, memory-carrying derivative of order α<1. In the memoryless case α=1, activity grows as sqrt(t) early and t later, with a characteristic crossover time. With memory, numerical solution shows effective activity declining with node age, and degree distributions broadening so late nodes can also become hubs. The claimed consequence is that memory redistributes hub potential and makes network evolution depend on memory length.","feed_headline":"Memory makes old network hubs lose their pull","feed_subtitle":"Fractional memory slows node activity after an early peak, redistributing who becomes a hub.","key_machinery":"The central object is a Caputo-style fractional differential equation and its discrete predictor-corrector solution. The paper's memory mechanism is a power-law kernel κ(t−t')=(t−t')^{α−2}/Γ(α−1), inserted into the rate equation so that current node activity is a weighted convolution of all past activities, with the fractional order α controlling how far back memory reaches. The predictor-corrector scheme with coefficients b_n=((n+1)^α−n^α)/Γ(α+1) supplies the aging weights that make older events contribute less, and the numerical solution of that scheme produces the reported rise-then-fall activity curves.","core_discovery":"The central discovery is an aging effect: when link formation retains a power-law memory of past activity, the effective node activity k̄_i(t) reaches a maximum and then declines, so older nodes lose their attractiveness and receive fewer new connections. This is presented as a general property of memory in temporal networks, controlled by fractional order α: α=1 reproduces the ordinary preferential attachment growth, while smaller α (longer memory) lowers and delays the peak. Consequently, in memoryful networks high-degree nodes need not be early ones; the degree distribution becomes broader, and link weights and assortativity differ from the memoryless limit.","pith_inferences":["The same fractional-memory attachment rule, applied to citation or recommender networks, would suppress rich-get-richer concentration and increase turnover among top nodes; the paper does not draw this implication.","One could fit α from real interaction logs by matching the empirically observed peak time of node activity to the discrete b_n solution, yielding a direct estimate of network memory length.","Because the memoryless model already has a crossover time, dense event streams should reveal aging effects earlier and more sharply than sparse streams, a testable prediction for empirical temporal networks."],"forward_implications":["In the memoryless limit α=1, node activity grows as $\\sqrt{t}$ at early times and as $t$ later, with a crossover at $t^* = c^2 N^2 / (4m^2)$.","For memory orders α<1, effective activity rises to a peak and then declines, so a node's probability of receiving new links falls as the node ages.","Longer memory (smaller α) reduces and delays the peak activity, slowing the growth of active nodes relative to the memoryless case.","Memory broadens the degree distribution away from the Gaussian form of the memoryless model, so nodes that join later can also become hubs.","Dense temporal networks exhibit the characteristic crossover and memory effects more clearly than sparse ones."],"supporting_citations":[{"why":"Supplies the mean-field solution and the sqrt(t)-versus-t growth regimes for the memoryless preferential attachment model that the paper extends.","marker":"[42]"},{"why":"Introduces the Caputo fractional derivative formalism used to encode memory in the activity equation.","marker":"[34]"},{"why":"Provides the predictor-corrector algorithm used to compute the effective node activity in the discrete scheme.","marker":"[43]"},{"why":"Gives the stability and implementation basis for the predictor-corrector numerical solution.","marker":"[44]"},{"why":"Introduces the preferential attachment mechanism that defines the memoryless baseline model.","marker":"[41]"}],"fun_headline_variants":["Memory in networks makes hubs age out of favor","Old nodes lose charm when network remembers past activity","Fractional memory pushes networks to favor younger nodes","Aging hubs decline when links have memory","Network memory lets younger nodes take over as hubs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire aging prediction depends on the claim that inserting a power-law memory kernel into the activity equation yields a valid fractional differential equation of the Caputo form; if that step is not mathematically well-defined, the rise-and-fall of node activity is not a consequence of the model.","fun_headline_variants_meta":{"raw":{"variants":["Memory in networks makes hubs age out of favor","Old nodes lose charm when network remembers past activity","Fractional memory pushes networks to favor younger nodes","Aging hubs decline when links have memory","Network memory lets younger nodes take over as hubs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000639,"raw_usage":{"total_tokens":2878,"prompt_tokens":816,"completion_tokens":2062,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":1992}},"tokens_in":432,"tokens_out":2062,"duration_ms":15777,"temperature":1.0,"reasoning_tokens":1992,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:57:24.432973+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check whether the kernel (t−t')^{α−2}/Γ(α−1) is integrable on [t0,t] for α<1, or solve the same preferential-attachment equation using the standard Caputo derivative (with the first derivative of node activity inside the integrand) for α=0.5 and compare the resulting activity curve to the paper's Fig. 3.","supporting_citations":[{"cited_title":"Barab´asi, R","cited_arxiv_id":null,"evidence_quote":"Supplies the mean-field solution and the sqrt(t)-versus-t growth regimes for the memoryless preferential attachment model that the paper extends."},{"cited_title":"Diethelm, N","cited_arxiv_id":null,"evidence_quote":"Provides the predictor-corrector algorithm used to compute the effective node activity in the discrete scheme."},{"cited_title":"Garrappa, On linear stability of predictor–corrector algorithms for fractional di fferential equations, International Journal of Computer Mathematics 87 (10) (2010) 2281–2290","cited_arxiv_id":null,"evidence_quote":"Gives the stability and implementation basis for the predictor-corrector numerical solution."},{"cited_title":"Barab ´asi, R","cited_arxiv_id":null,"evidence_quote":"Introduces the preferential attachment mechanism that defines the memoryless baseline model."}],"review_version":1}