{"id":"7743e252-cd39-42ec-9560-3db1f298405a","arxiv_id":"1908.07066","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For random threshold graphs under a general fitness scaling, the empirical fraction of nodes with degree d converges only in distribution to a non-degenerate random variable, not in probability to the nodal degree pmf.","lead":"Random threshold graphs with exponentially distributed fitness have a nodal degree distribution that converges to a nice limit, but the network-wide empirical degree distribution stays randomly fluctuating instead of converging to it. This paper proves the fluctuation persists for a broad class of fitness distributions, undermining earlier claims that such graphs give a scale-free alternative to preferential attachment models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1's non-degeneracy claim Var[Π(d)]>0 is asserted but not proved under general Assumption 1; the cited Section 7 argument covers only the constant-λ case and defers the exponential case to earlier references.","rationale":"The reader's weakest assumption correctly identifies the non-degeneracy gap, and my independent reading confirms this is the load-bearing issue. The method-of-moments and characteristic-function machinery is otherwise sound: boundedness of Pn(d) justifies the moment series, Proposition 3.4 supplies the moment limits, and the Cramér-Lévy argument yields weak convergence once the limit function is known to be a characteristic function. The conditional-Poisson representation in Sections 8-10 is coherent, and the constant-λ and exponential examples are consistent with the claimed behavior. However, the paper does not provide a proof that Var[Π(d)]>0 for every d under the full Assumption 1; the cited 'discussion at the end of Section 7' covers only constant λ and refers to unpublished or separate references for the exponential case. Because this positivity is precisely what converts distributional convergence into non-convergence in probability, the central applied claim remains conditional. The reader's CONDITIONAL verdict is therefore appropriate, and the proposed analytical check would settle whether the gap is a mere omission or a genuine failure.","tokens_in":18193,"tokens_out":21744,"duration_ms":219247,"concrete_test":"Using the r=2 pgf (49), derive the conditional diagonal probability: for Λ1,Λ2 iid with Λ=λ(ξ), P(D1=D2=d | Λ1=a, Λ2=b) = e^{-max(a,b)} min(a,b)^d / d!, hence Var[Π(d)] = E[e^{-max(Λ1,Λ2)} min(Λ1,Λ2)^d]/d! - (E[e^{-Λ}Λ^d]/d!)^2. Verify this inequality analytically or by high-precision quadrature for a representative nonconstant λ satisfying Assumption 1, e.g., λ(x)=e^{αx} for exponential fitness and a two-point mixture of Λ, for d=0,1,2. If a counterexample is found, Theorem 3.1 is false; if positivity holds, the proof gap in Section 7 can be closed by inserting this derivation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires Var[Π(d)]>0 for every d under the full Assumption 1, because only this positivity turns the weak convergence Pn(d;θ*_n) => Π(d) into the advertised failure of (7). Section 5 reduces this to Var[Π(d)] = P[D1=d,D2=d] - P[D=d]^2 in (39), but the proof of positivity is deferred to 'the discussion at the end of Section 7.' That discussion establishes non-independence of D1,...,Dr, then treats only the constant-λ case (where Π(d) is two-point and the variance is explicit) and the exponential case, which is cited to references [14,16]. No argument is supplied for a nonconstant λ satisfying Assumption 1 outside the exponential case. Non-independence of D1,D2 does not by itself imply positive covariance of every pair of indicators 1[D1=d], 1[D2=d], so the asserted inequality could in principle fail for some d. Since all downstream conclusions — failure of (7) and rejection of the scale-free interpretation — are loaded on this positivity, the theorem as stated is conditional on a missing proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies random threshold graphs with i.i.d. nonnegative fitness variables under Assumption 1, a tail-scaling condition on the fitness distribution. Under this assumption the authors prove that a single node's degree converges in distribution to a conditionally Poisson limit D, and they establish that the empirical degree fraction P_n(d; θ*_n) converges weakly to a [0,1]-valued limit Π(d) whose moments are given by P(D_1 = d, ..., D_r = d). The advertised conclusion is that Π(d) is non-degenerate, so the empirical degree fraction does not converge in probability to the nodal degree pmf; in the exponential case this is used to argue that random threshold graphs do not provide a scale-free alternative to the Barabási-Albert model. The proof proceeds through a multivariate pgf convergence result (Proposition 3.3), a method-of-moments step (Proposition 3.4), and identification of the limit via its characteristic function. Simulations for the exponential case illustrate the claimed non-degeneracy.","tokens_in":18358,"tokens_out":5139,"duration_ms":51672,"significance":"If the non-degeneracy claim is fully established, the paper makes a conceptually important point: in a homogeneous random graph model, the network-wide empirical degree distribution and the single-node degree distribution can have different asymptotic behavior. This has direct implications for empirical work that uses the empirical degree distribution as a proxy for the nodal distribution, and it corrects a claim in the threshold-graph literature about scale-free alternatives to preferential attachment. The proof strategy is generally rigorous and self-contained: the joint pgf convergence and the moment computations are carried out in detail, no fitted parameters appear, and the simulations support the qualitative conclusion. The main caveat is that the proof of Var[Π(d)] > 0, which is load-bearing for the headline conclusion, is deferred and not supplied under the full generality of Assumption 1.","major_comments":[{"comment":"The assertion Var[Π(d)] > 0 in Theorem 3.1 is not proved under the full Assumption 1. Section 5 refers to 'the discussion at the end of Section 7', but that discussion treats only the constant-λ case, where Π(d) is two-point with an explicit variance, and the exponential case, which is deferred to references [14,16]. Non-independence of D1,...,Dr, even if fully established, does not imply Cov(1[D1=d], 1[D2=d]) > 0 for every fixed d. Since Corollary 3.2 and the advertised failure of (7) depend precisely on this positivity, Theorem 3.1 as stated is conditional on a missing proof. Please supply a direct argument that Var[Π(d)] > 0 for every d under Assumption 1, or state and prove a weaker theorem.","section":"Section 5, Eq. (39); end of Section 7"},{"comment":"The inequality (50), asserting that the joint pgf of (D1,D2) differs from the product of the marginal pgfs, is stated after 'comparing (48) and (49)' but no proof is given. This is part of the claimed non-independence in Proposition 3.3 and is also the only qualitative support for the variance claim. For the constant-λ case the computation is explicit, but for a general nonconstant λ satisfying Assumption 1 a derivation is needed. The same comparison must also yield positive covariance of the indicators 1[D1=d] and 1[D2=d] for each d, which is a stronger statement than mere non-independence.","section":"Section 7, Eq. (50)"}],"minor_comments":[{"comment":"The text contains a typo: 'high probbability' should be 'high probability'.","section":"Section 4"},{"comment":"The phrase 'The rv Π(d) is non-degenerate Fix d = 0, 1,...' appears to have a missing line break or punctuation; it should read as a subsection heading or a new sentence.","section":"Section 4"},{"comment":"In the constant-λ display, 'd,d ′ = 01, 2,...' should read 'd,d′ = 0,1,2,...'.","section":"Section 7"},{"comment":"The abstract and introduction state the failure of convergence in probability as a conclusion; given the missing variance proof, the wording should be softened or the proof supplied.","section":"Introduction"},{"comment":"It would be helpful to state explicitly that Assumption 1 is a regular-variation-type condition that many continuous distributions do not satisfy, so the theorem applies only to fitness distributions admitting such a scaling.","section":"Assumption 1"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of math.PR and the central idea is valuable. The missing variance proof is likely repairable, so I do not recommend rejection. However, the reliance on references [14,16] for a key step in the exponential case, and the complete absence of an argument for the general case, make this a load-bearing gap that must be fixed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. This is a good paper and worth taking seriously. The central result, Theorem 3.1, is exactly the right way to settle the question the authors pose: under Assumption 1, the empirical degree fraction P_n(d; theta*_n) converges in distribution to a random variable Pi(d), not in probability to the nodal pmf. That corrects an over-reading of the scale-free threshold graph literature and gives a clean explanation of why smoothed empirical degree distributions can look like p_Fuj(d) while individual realizations do not. The proof strategy, built on joint pgf convergence, the method of moments, and characteristic function identification, is sound in outline and, except for one point, carefully executed. The simulations are simple but genuinely illustrate the point. Assumption 1 is honestly stated, and the paper gives real examples, exponential and Pareto, where it holds. The self-citations also seem legitimate: the exponential case was established in earlier work, and the present paper extends it, so citing those results is not a red flag.\n\nThe soft spot is the non-degeneracy claim Var[Pi(d)] > 0 for all d under general Assumption 1. Section 5 refers to the discussion at the end of Section 7, but that discussion only establishes that D_1,...,D_r are not independent, then works out the constant-lambda case and cites the exponential case back to references [14,16]. Non-independence of the vector does not imply positive covariance for every pair of indicators 1[D_1=d], 1[D_2=d], so as written, Theorem 3.1 and Corollary 3.2 are conditional on a missing argument. I would be surprised if the claim is false; the natural missing step is some form of positive association. But it is not in the preprint. For the exponential case, where the variance is handled in the earlier references, the main message about the scale-free interpretation is safe. For the general statement, the gap needs to be addressed.\n\nWho is this for: people working on threshold graph asymptotics, and anyone citing Caldarelli et al. as evidence for scale-free threshold graphs. It deserves peer review; the fix is likely short, but a referee should insist on seeing it. I would bring it to a reading group.","headline":"A mostly rigorous, well-written paper showing that empirical degree fractions in random threshold graphs converge only in distribution to a non-degenerate limit; the one real gap is an unproved positivity claim under the full Assumption 1, which should be fixed before publication.","tokens_in":18943,"tokens_out":3845,"would_cite":true,"duration_ms":40989,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C80","60F05","60G70"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under one scaling assumption, the fraction of nodes of a given degree in a random threshold graph converges to a non-degenerate random variable, not to the nodal degree distribution.","keywords":["random threshold graphs","degree distribution","empirical degree distribution","nodal degree distribution","fitness model","scale-free networks","weak convergence","characteristic function"],"falsifier":"For exponentially distributed fitness with $\\lambda=1$ and $\\theta^\\star_n=\\log n$, compute the variance $\\mathrm{Var}[\\Pi(d)]=\\mathrm{Cov}(\\mathbf{1}[D_1=d],\\mathbf{1}[D_2=d])$ from the joint generating function (49); if for some $d$ the variance vanishes, the claimed non-degeneracy fails. Equivalently, simulate many independent graphs at large $n$ and check whether the histogram of $P_n(d;\\log n)$ collapses to a point at $p_{\\rm Fuj}(d)$ rather than spreading.","tokens_in":17927,"feed_emoji":"📊","tokens_out":4679,"duration_ms":46520,"temperature":0.7,"pith_summary":"The paper asks whether the fraction of nodes of a given degree in a random threshold graph settles to the same limiting distribution as the degree of a single node. It shows the answer is no: under a broad scaling assumption on the fitness distribution, for each degree $d$ the empirical fraction $P_n(d;\\theta^\\star_n)$ converges only in distribution, to a non-degenerate random variable $\\Pi(d)$, so the large-network histogram never settles down. In the exponential case this overturns the idea that these graphs provide a scale-free alternative to the Barabási–Albert model. The distribution of $\\Pi(d)$ is identified through its characteristic function, built from joint degree probabilities of exchangeable limiting degrees.","feed_headline":"Degree counts in random threshold graphs never settle down","feed_subtitle":"Even with exponential fitness, the empirical degree distribution stays random, so it cannot replace the nodal degree law.","key_machinery":"The argument rests on Assumption 1, which says the fitness distribution $F$ admits a threshold scaling $\\theta^\\star_n\\to\\infty$ for which $n(1-F(\\theta^\\star_n-x))$ converges to a non-identically-zero function $\\lambda(x)$. This scaling makes the single-node degree converge to a conditionally Poisson variable $D$ with pmf $P[D=d]=E[\\lambda(\\xi)^d e^{-\\lambda(\\xi)}/d!]$. The main technical step, Proposition 3.3, extends this to the joint limit of $r$ node degrees via their multivariate probability generating function, expressed through order statistics and the random permutation arranging the fitness values. The method of moments then yields the characteristic function of $\\Pi(d)$ from the joint probabilities $P[D_1=d,\\ldots,D_r=d]$, and exchangeability of the limiting degrees supplies the factorial moments.","core_discovery":"The paper's central result, Theorem 3.1, states that under Assumption 1 there exists, for each $d=0,1,\\ldots$, a non-degenerate $[0,1]$-valued random variable $\\Pi(d)$ such that $P_n(d;\\theta^\\star_n)$ converges in distribution to $\\Pi(d)$, with $E[\\Pi(d)]=P[D=d]$ and $\\mathrm{Var}[\\Pi(d)]>0$. Consequently the empirical degree fraction does not converge in probability to the nodal degree pmf, so the customary equation $P_n(d;\\theta^\\star_n)\\to p_{\\rm Fuj}(d)$ fails. The distribution of $\\Pi(d)$ is given only through its characteristic function\n$$\\Phi_d(t)=1+\\sum_{r=1}^\\infty \\frac{(it)^r}{r!}\\,P[D_1=d,\\ldots,D_r=d],$$\nwhere $(D_1,\\ldots,D_r)$ are the exchangeable but dependent limiting degrees of $r$ distinct nodes. In the exponential case this directly contradicts the claim that random threshold graphs provide a scale-free alternative to the Barabási–Albert model; the two models cannot be compared through their empirical degree distributions.","pith_inferences":["A testable extension: for other homogeneous random graph models built from i.i.d. latent variables through a smooth edge kernel, the same random-limit phenomenon should appear whenever a Poisson-type scaling exists; plotting $\\mathrm{Var}[P_n(d;\\theta_n)]$ against $n$ would reveal whether the variance plateaus at a positive value.","If a closed-form expression for $\\mathrm{Var}[\\Pi(d)]$ can be derived for the exponential case, it would give practitioners a direct confidence interval for histogram fluctuations; the paper leaves this as an open computation.","The result implicitly challenges the common data-analysis practice of reading scale-freeness from a single observed degree histogram: even the correct homogeneous model would show run-to-run fluctuations that averaging over nodes within one network cannot remove."],"forward_implications":["For exponentially distributed fitness with $\\theta^\\star_n=\\lambda^{-1}\\log n$, the fraction of nodes of degree $d$ does not settle to a fixed value as $n$ grows; sample histograms fluctuate by an amount that does not vanish.","The empirical degree distribution of a single large random threshold graph cannot be used as an estimator or proxy for the limiting nodal degree pmf $p_{\\rm Fuj}$.","Even in a homogeneous random graph model, the network-wide empirical degree distribution and the single-node degree distribution can carry fundamentally different information.","Random threshold graphs with exponential fitness do not provide a scale-free alternative to the Barabási–Albert model as claimed in earlier work; the two models cannot be meaningfully compared in terms of their degree distributions.","The same random-limit behavior holds for every fitness distribution satisfying Assumption 1, so the phenomenon is not an artifact of the exponential case."],"supporting_citations":[{"why":"Supplies the exponential-case convergence of the single-node degree to the power-law pmf $p_{\\rm Fuj}$ that the paper's empirical-degree result overturns.","marker":"[10]"},{"why":"Proposes the fitness-based random threshold graph and claims scale-free behavior, the claim this paper directly addresses.","marker":"[4]"},{"why":"Introduces the Barabási–Albert model and the power-law scaling that motivates the comparison.","marker":"[1]"},{"why":"Proves the empirical degree distribution of the Barabási–Albert model converges in probability to $p_{\\rm BA}$, the contrasting benchmark.","marker":"[3]"},{"why":"Extends the fitness argument to arbitrary scale-free networks and is another target of the paper's no-comparison conclusion.","marker":"[17]"},{"why":"Earlier counterexample for exponential fitness showing failure of convergence in probability, generalized here under Assumption 1.","marker":"[16]"}],"fun_headline_variants":["Threshold graph degree fractions stay forever random","Empirical degree counts never converge in random threshold graphs","Random threshold graphs: degree distribution is a random variable","Degree counts in threshold graphs converge only in distribution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result is conditional on Assumption 1: the fitness distribution must admit a threshold scaling $\\theta^\\star_n\\to\\infty$ for which $n(1-F(\\theta^\\star_n-x))$ converges to a non-identically-zero $\\lambda(x)$; distributions without such regular variation are outside the scope, and even under the assumption the paper does not supply a proof of the asserted strict positivity of $\\mathrm{Var}[\\Pi(d)]$ for every $d$.","fun_headline_variants_meta":{"raw":{"variants":["Threshold graph degree fractions stay forever random","Empirical degree counts never converge in random threshold graphs","Random threshold graphs: degree distribution is a random variable","Degree counts in threshold graphs converge only in distribution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1550,"prompt_tokens":1028,"completion_tokens":522,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":463}},"tokens_in":644,"tokens_out":522,"duration_ms":5854,"temperature":1.0,"reasoning_tokens":463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:27:20.467064+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For exponentially distributed fitness with $\\lambda=1$ and $\\theta^\\star_n=\\log n$, compute the variance $\\mathrm{Var}[\\Pi(d)]=\\mathrm{Cov}(\\mathbf{1}[D_1=d],\\mathbf{1}[D_2=d])$ from the joint generating function (49); if for some $d$ the variance vanishes, the claimed non-degeneracy fails. Equivalently, simulate many independent graphs at large $n$ and check whether the histogram of $P_n(d;\\log n)$ collapses to a point at $p_{\\rm Fuj}(d)$ rather than spreading.","supporting_citations":[{"cited_title":"Limit theorems for the average distance and the degree distribution of the threshold network model,","cited_arxiv_id":null,"evidence_quote":"Supplies the exponential-case convergence of the single-node degree to the power-law pmf $p_{\\rm Fuj}$ that the paper's empirical-degree result overturns."},{"cited_title":"Scale-free networks from varying vertex intrinsic fitness,","cited_arxiv_id":null,"evidence_quote":"Proposes the fitness-based random threshold graph and claims scale-free behavior, the claim this paper directly addresses."},{"cited_title":"Emergence of scaling in random networks,","cited_arxiv_id":null,"evidence_quote":"Introduces the Barabási–Albert model and the power-law scaling that motivates the comparison."},{"cited_title":"The degree sequence of a scale free random graph process,","cited_arxiv_id":null,"evidence_quote":"Proves the empirical degree distribution of the Barabási–Albert model converges in probability to $p_{\\rm BA}$, the contrasting benchmark."},{"cited_title":"Vertex intrinsic fitness: How to produce arbitrary scale-free networks,","cited_arxiv_id":null,"evidence_quote":"Extends the fitness argument to arbitrary scale-free networks and is another target of the paper's no-comparison conclusion."},{"cited_title":"Asymptotic degree distributions in large (homogeneous) random networks: A little theory and a counterexample","cited_arxiv_id":"1710.11064","evidence_quote":"Earlier counterexample for exponential fitness showing failure of convergence in probability, generalized here under Assumption 1."}],"review_version":1}