{"id":"5142f244-dd4f-4fa2-9b95-e193b632cf4b","arxiv_id":"2509.05650","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For the stationary branching process X = A + sum_{i=1}^X B_i with index-one heavy tails and E[B]=b<1, the tail is asymptotic to 1/((1-b)(1+x)).","lead":"The paper derives the exact large-value probability for a branching population with heavy-tailed immigration and reproduction, showing the tail is amplified by the factor 1/(1-b). It also gives a smaller logarithmic correction term.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3.3's proof is invalid: Step 4's uniform one-big-jump fails near k=x/E[Y]; the paper does not supply the standard repair, so Proposition 3.5 and the refined theorem are not rigorously derived.","rationale":"The reader is right that Lemma 3.3 is the weakest point and that Step 4's uniform assertion is false; the example with k near x/E[Y] shows P(S_k>x) can be bounded away from zero even when kF(x)→0. However, the reader's stronger statement that Lemma 3.3 itself is false may be too hasty: for a fixed counting variable N, the probability mass in the problematic transition window is generally negligible relative to P(N>x/m), and the lemma is a known type of random-sum tail asymptotic. The paper's proof is nevertheless unsound, and because Proposition 3.5, Lemma 3.2, and the summation in Corollary 3.10 all rely on this lemma or on the same uniform WLLN assertion, the central refined asymptotic is not rigorously derived as written. The result may be repairable under the paper's specific assumptions, but the current derivation requires a corrected proof or a citation to a standard theorem. This matches the reader's CONDITIONAL verdict, so no change is needed.","tokens_in":12838,"tokens_out":40514,"duration_ms":393229,"concrete_test":"Analytically verify Lemma 3.3 in the paper's exact regime: let F have tail cL(x)/x, E[Y]=m, and let N have tail (1+x)^{-1}; derive P(Σ_{i=1}^N Y_i>x) by splitting at k=(1−ε)x/m and showing the transition window |k−x/m|≤kL(k) contributes o(P(N>x/m)). If this bound fails, Proposition 3.5 has a missing term of the same order as P(A>xb^{-n}); if it holds, replace the faulty Step 4 with this standard argument and check that the subsequent summation over n remains valid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Lemma 3.3 (Uniform random-sum tail), used in Proposition 3.5 and Lemma 3.2. Its proof Step 4 asserts P(S_k>x)=kF(x)(1+o(1)) uniformly for h(x)<k≤x/m. This is false even for finite-mean subexponential Y with xF(x)→0: take k=x/m−1; then x−km=m, and P(S_k>x) tends to a positive constant (the centered sum has scale o(k) and the threshold shift is O(1)), while kF(x)→0. Lemma 3.2 Step 5 has the same defect at k just above x/m, where WLLN gives only P(S_k>x)→1/2, not →1. The final lemma may be true for a fixed counting variable N—the mass of N in the transition window is usually o(P(N>x/m))—but that is not proved or cited. Since Proposition 3.5 and hence Theorem 3.8/Corollary 3.10 rest on Lemma 3.3, the refined two-scale claim is not established as written. A correct proof needs a split at (1−ε)x/m plus a bound on the near-threshold window; the paper provides neither.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the stationary solution X of the branching fixed-point equation X = A + sum_{i=1}^X B_i under the assumptions P(A>x) ~ 1/(1+x), P(B>x) = L(x)/(1+x) with L(x) ~ (log x)^{-1-epsilon} for some epsilon>0, and b = E[B] in (0,1). The main result, Theorem 1.1, claims P(X>x) ~ 1/((1-b)(1+x)) plus an explicit logarithmic correction of order L(x) log x / (1+x), which is asymptotically negligible. The proof develops generation-level tail asymptotics for branching aggregates D_n, a cluster expansion of X, a uniform random-sum tail lemma, and a countable subexponential sum lemma. The central claims are Theorem 3.8 and Corollaries 3.9 and 3.10.","tokens_in":1643,"tokens_out":1853,"duration_ms":205893,"significance":"If the main theorem is correct, this is a meaningful refinement of known results for heavy-tailed branching processes with immigration: it sharpens the standard (1-b)^{-1} amplification of the immigration tail, gives a generation-by-generation decomposition, and quantifies an explicit but negligible logarithmic correction. The proof strategy is direct, parameter-free, and does not rely on fitting or circular reasoning; the cluster expansion is a useful structural representation. However, the central technical lemmas contain substantial gaps, and the proof as written is not rigorous. In particular, Lemma 3.3 is false as stated, and Lemma 3.2 and Lemma 3.7 have unjustified uniform or scaling steps. The main theorem may be salvageable, but a real revision is required.","major_comments":[{"comment":"Lemma 3.3 is false as stated, and the proof rests on a false uniform one-big-jump assertion. Take Y lognormal (subexponential with finite mean m) and N deterministic equal to floor(x/m). Then the right-hand side of the lemma is N P(Y>x) + 0 = o(1) because finite mean implies x P(Y>x) -> 0, while by the CLT P(S_N > x) with x ~ N m tends to 1/2. Step 4 asserts that P(S_k > x) = k P(Y>x)(1+o(1)) uniformly for h(x)<k<=x/m; at k=x/m the left side is bounded away from 0 while k P(Y>x) -> 0. Since Proposition 3.5 and Theorem 3.8 use Lemma 3.3, this invalidates the proof of the main theorem. A corrected lemma needs either extra hypotheses (for example, uniform decay of k P(Y>x) on the whole range and control of the transition window) or a split at (1-epsilon)x/m with an explicit estimate for the remaining window.","section":"Lemma 3.3, proof Step 4"},{"comment":"The assertion that P(sum_{i=1}^k D_{n,i} > x) -> 1 uniformly for all k > k_x = floor(x/m_n) is false. For k = k_x + O(1), the mean k m_n is x + O(1), and since D_n has finite variance in the setting of this paper (the tail is n b^{n-1} L(x)/x with L(x)=(log x)^{-1-epsilon} integrable), the CLT gives a limit of 1/2, not 1. This invalidates the derivation of S_2(x) ~ P(B > k_x). The induction may still be correct because the B-probability mass in an O(1) window around k_x is o(P(B>k_x)) for a smooth regularly varying B-tail, but that estimate is not proved or cited and must be supplied.","section":"Lemma 3.2, proof Step 5"},{"comment":"The proof uses the equivalence F_k(x/2) ~ F_k(x) for subexponential F_k, but subexponentiality only gives long-tailedness, i.e., F(x+y) ~ F(x) for fixed y; it does not imply F(x/2) ~ F(x). For Weibull-type subexponential tails, F(x/2)/F(x) can grow exponentially. Consequently the displayed comparison sum_{k>K} F_k(x/2) ~ sum_{k>K} F_k(x) is not justified, and the upper bound in Step 2 does not follow from the stated assumptions. Since Lemma 3.7 is used in Theorem 3.8 Step 5 to pass from finite to countable sums, the lemma needs either additional regularly-varying assumptions (which are satisfied by the Y_n in this paper) or a different proof.","section":"Lemma 3.7, proof Step 2"}],"minor_comments":[{"comment":"The proof is internally inconsistent about the meaning of F: Lemma 3.1 defines F as the tail, but Steps 4 and 6 use F(h(x)) as a distribution function. As written, Step 4 gives I_1(x) ~ \\bar{F}(h(x)) \\bar{F}(x) = o(\\bar{F}(x)), while Step 6 asserts \\bar{F}(h(x)) \\uparrow 1, which is contradictory. The proof should consistently distinguish the distribution function from its tail.","section":"Lemma 3.1, proof"},{"comment":"The phrase 'by dominated convergence' is used to sum asymptotic equivalences over n, but the asymptotic P(A > x b^{-n}) ~ b^n/x is only stated for each fixed n. Justifying the series passage requires a finite-split or uniformity argument.","section":"Lemma 3.6, proof Part I"},{"comment":"The reference to 'Lemma 1 for alpha=1 tails' is incorrect; the intended reference appears to be Lemma 3.1.","section":"Proposition 3.5"},{"comment":"References [4], [5], and [13] are listed but never cited in the body; please cite them or remove them.","section":"References"},{"comment":"The display in Step 2 of the proof of Lemma 3.4 is garbled and does not clearly show the rearrangement of sums; the unfolding algebra should be rewritten carefully.","section":"Lemma 3.4, proof Step 2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: I believe the main theorem is likely true and repairable, but the current proof is not reliable. Lemma 3.3 is false as stated, and Lemma 3.2 Step 5 and Lemma 3.7 are not rigorously justified. The manuscript also contains garbled displays and uncited references, suggesting it was assembled quickly. I do not see grounds for rejection on novelty or circularity grounds; the two-scale refinement is a genuine contribution if properly proved. However, the revision will require real mathematical work, not just exposition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper claims a refined tail asymptotic for a heavy-tailed branching process with immigration, adding an explicit logarithmic correction at the boundary index-one case. That result is likely true and is a natural extension of Foss and Miyazawa, and the cluster expansion is a genuinely tidy way to decompose the fixed point. But the proof has load-bearing errors in the random-sum lemmas, so the derivation as written is not sound.\n\nWhat is new: the boundary case with L(x) ~ (log x)^{-1-eps}, the explicit second-order term of order L(x) log x / (1+x), and the cluster expansion representation. The cluster expansion is a nice structural device, and the paper is clearly organized. No fitted parameters, no self-citation issues, no invented entities.\n\nThe soft spots are concentrated in Lemmas 3.1–3.3. Lemma 3.3 is false as stated. Take N deterministic with N ≈ x/E[Y] and Y with a Pareto(2) tail. Then the left side tends to about 1/2, while the right side tends to 0. The proof's Step 4 asserts uniform one-big-jump behavior for all k up to x/m without requiring k F(x) → 0, which fails near k = x/m. Lemma 3.2's Step 5 has the same defect: for k just above x/m, the sum's mean is barely above x and the exceedance probability is about 1/2, not 1, so the claimed uniform convergence to 1 is false. Lemma 3.1's proof also confuses the tail F(h) with the cdf, writing I1 ~ F(h) F(x) where it should be (1-F(h)) F(x); the lemma itself is true (regularly varying tails are subexponential), but the proof as written is invalid.\n\nThat said, the main theorem may well be correct. In the intended application, the missing condition x F_{D_n}(x) → 0 does hold because L(x) → 0, so the contested near-threshold terms are negligible. A standard repair—splitting at (1-ε)x/m and bounding the window—would likely fix the argument. But that repair is not in the paper, and Proposition 3.5 and Theorem 3.8 rest directly on the faulty lemmas.\n\nMy bottom line: this is a conditional accept at best, and only after the author fixes the random-sum lemmas and the cdf/tail slip. The cluster expansion and the explicit correction term make it worth a serious referee, not a desk reject. For a reading group, it is a useful case study in how uniform one-big-jump claims can fail near the fluid threshold, but I would not cite it in its current form.","headline":"Plausible refinement of a known tail asymptotic, but the proof as written rests on a false random-sum lemma and an erroneous estimate near the fluid threshold; worth refereeing but needs major revision.","tokens_in":13587,"tokens_out":5035,"would_cite":false,"duration_ms":45186,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J80","60G70","60K25","60F10","60E05","60K05","60G50"],"pacs":[],"model":"deepseek-v4-flash","headline":"At boundary heavy-tail index one, the stationary tail of a branching process with immigration is the immigration tail scaled by $1/(1-b)$, plus an explicit smaller logarithmic term.","keywords":["branching processes","immigration","fixed-point equations","heavy-tailed distributions","subexponentiality","principle of a single big jump","tail asymptotics","cluster expansion"],"falsifier":"Take $A$ with $P(A>x)=1/(1+x)$ and $B$ with $P(B>x)=1/[(1+x)(\\log(e+x))^2]$, so $L(x)\\sim(\\log x)^{-1}$; computing the fixed-point tail numerically should show whether the difference from the displayed expansion is $o(1/x)$. A sharper check targets the random-sum lemma itself: for $Y$ with tail $1/(1+x)$ and deterministic $N=\\lfloor x/\\mathbb{E}[Y]\\rfloor$, the left side stays near $1/2$ while $kP(Y>x)\\to 0$, so any proof using that uniform step without extra conditions would predict the wrong value.","tokens_in":12617,"feed_emoji":"🌳","tokens_out":7926,"duration_ms":67359,"temperature":0.7,"pith_summary":"The paper studies the stationary solution $X$ of the branching-and-immigration fixed-point equation $X \\stackrel{d}{=} A + \\sum_{i=1}^{X} B_i$, where $A$ is immigration and $B$ is offspring. It assumes both are heavy-tailed at the boundary index one: $P(A>x)\\sim(1+x)^{-1}$ and $P(B>x)=L(x)/(1+x)$ with $L(x)\\sim(\\log x)^{-1-\\varepsilon}$. The main theorem proves $P(X>x)\\sim\\frac{1}{(1-b)(1+x)}$ and refines this by identifying a second-scale correction of order $L(x)\\log x/(1+x)$, which is $o(1/(1+x))$. A sympathetic reader should care because the boundary case is where the one-big-jump heuristic is most delicate: the paper shows the tail still closes, and the correction records how logarithmic factors propagate generation by generation.","feed_headline":"One big jump sets the stationary tail of branching with immigration","feed_subtitle":"Immigration tail amplified by 1/(1-b), with an explicit but negligible logarithmic correction.","key_machinery":"The load-bearing construction is the cluster expansion $$X \\stackrel{d}{=} A + \\sum_{n\\ge 1}\\sum_{i=1}^{A_{n+1}} D_{n,i},$$ where $D_1\\stackrel{d}{=}B$ and $D_{n+1}\\stackrel{d}{=}\\sum_{j=1}^{B}D_{n,j}$. This writes the stationary solution as an independent superposition of immigration-weighted generation clusters. The proof combines generation-tail asymptotics $P(D_n>x)\\sim n b^{n-1}P(B>x)$, a uniform random-sum tail lemma for subexponential summands, a tail-summability lemma, and a countable closure principle for subexponential sums; together these transfer the single-big-jump asymptotics from one sum to the infinite expansion.","core_discovery":"The central claim is that, under the stated conditions, the tail of the stationary solution obeys $$P(X>x) \\sim \\frac{1}{(1-b)(1+x)} + \\left(\\sum_{n\\ge 1} n $b^{{n-1}}$ \\log(x $b^{{-n}}$)\\right)\\frac{L(x)}{1+x}\\,(1+o(1)),\\quad x\\to\\infty.$$ The first term is the immigration tail amplified by the factor $(1-b)^{-1}$; the second term is asymptotically negligible but explicit, with the series encoding how logarithmic corrections in the offspring tail accumulate over generations. This is a boundary version of the principle of a single big jump: at index one the extreme events are still driven by one large contribution, but their asymptotics require summing over the branching generations.","pith_inferences":["Because the theorem's dominant term is purely $A$-driven, the same leading tail should hold for any slowly varying $L$ that still satisfies the uniform one-big-jump control; the explicit second scale may change when $L(x)\\log x$ does not vanish.","The two-scale formula suggests a phase transition at the boundary: for $L(x)\\sim(\\log x)^{-1-\\varepsilon}$ the branching correction is negligible, while for heavier $L$ the correction would become comparable to or larger than the leading $1/(1+x)$ term.","The same generation-cluster decomposition should produce analogous refined asymptotics in continuous-state subordinator versions and second-order branching processes with immigration, since those models share the recursive fixed-point structure.","In queueing or risk terms, the factor $1/(1-b)$ gives an operational reading: the stationary tail is the single-immigration tail inflated by the expected number of feedback generations, and the logarithmic term tracks the slow variation of the offspring tail."],"forward_implications":["The dominant tail is $P(X>x)\\sim(1/(1-b))\\,P(A>x)$, so branching magnifies the immigration tail by exactly the mean total offspring factor $1/(1-b)$.","Generation aggregates satisfy $P(D_n>x)\\sim n b^{n-1}P(B>x)$, giving a concrete decomposition of extremes by depth in the branching tree.","The second-order correction is of order $L(x)\\log x/(1+x)$ and, because $L(x)\\sim(\\log x)^{-1-\\varepsilon}$, it is $o(1/(1+x))$; the stated asymptotic is genuinely two-scale.","Countably infinite subexponential sums inherit tail asymptotics equal to the sum of the individual tails whenever the far-tail mass is negligible, so recursive fixed-point solutions remain tractable.","The cluster expansion represents the stationary solution as an infinite sum of independent components, which the paper proposes could support simulation and numerical approximation of the tail."],"supporting_citations":[{"why":"Introduces the fixed-point model and heavy-tailed tail asymptotics for branching with state-independent immigration that this paper refines.","marker":"[14]"},{"why":"Supplies the standard definitions and closure facts for subexponential and long-tailed distributions used in the tail lemmas.","marker":"[6]"},{"why":"Establishes the one-big-jump asymptotics for maxima of heavy-tailed random walks that the random-sum lemma adapts to sums.","marker":"[7]"},{"why":"Provides the implicit renewal and tree-based power-tail framework that motivates the cluster expansion.","marker":"[9]"},{"why":"Gives the regular-variation fixed-point results for branching and queueing processes that the boundary-index-one result extends.","marker":"[11]"}],"fun_headline_variants":["Branching amplifies heavy-tail immigration by 1/(1-b)","Single big jump persists at boundary for branching with immigration","Refined tail: branching adds negligible log correction to immigration","Boundary index one: single big jump still drives the tail","Explicit log correction in branching tail at heavy-tail boundary"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on a uniform one-big-jump estimate: for sums of the subexponential variables involved, the tail must behave like $k$ times the single-variable tail uniformly for every $k$ up to about $x/\\mathbb{E}[Y]$, and that uniform control is not automatic from subexponentiality alone.","fun_headline_variants_meta":{"raw":{"variants":["Branching amplifies heavy-tail immigration by 1/(1-b)","Single big jump persists at boundary for branching with immigration","Refined tail: branching adds negligible log correction to immigration","Boundary index one: single big jump still drives the tail","Explicit log correction in branching tail at heavy-tail boundary"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001338,"raw_usage":{"total_tokens":5359,"prompt_tokens":785,"completion_tokens":4574,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":401,"completion_tokens_details":{"reasoning_tokens":4491}},"tokens_in":401,"tokens_out":4574,"duration_ms":27993,"temperature":1.0,"reasoning_tokens":4491,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:25:02.892283+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $A$ with $P(A>x)=1/(1+x)$ and $B$ with $P(B>x)=1/[(1+x)(\\log(e+x))^2]$, so $L(x)\\sim(\\log x)^{-1}$; computing the fixed-point tail numerically should show whether the difference from the displayed expansion is $o(1/x)$. A sharper check targets the random-sum lemma itself: for $Y$ with tail $1/(1+x)$ and deterministic $N=\\lfloor x/\\mathbb{E}[Y]\\rfloor$, the left side stays near $1/2$ while $kP(Y>x)\\to 0$, so any proof using that uniform step without extra conditions would predict the wrong value.","supporting_citations":[{"cited_title":"Foss and M","cited_arxiv_id":null,"evidence_quote":"Introduces the fixed-point model and heavy-tailed tail asymptotics for branching with state-independent immigration that this paper refines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the standard definitions and closure facts for subexponential and long-tailed distributions used in the tail lemmas."},{"cited_title":"Foss and S","cited_arxiv_id":null,"evidence_quote":"Establishes the one-big-jump asymptotics for maxima of heavy-tailed random walks that the random-sum lemma adapts to sums."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the implicit renewal and tree-based power-tail framework that motivates the cluster expansion."},{"cited_title":"Asmussen and S","cited_arxiv_id":null,"evidence_quote":"Gives the regular-variation fixed-point results for branching and queueing processes that the boundary-index-one result extends."}],"review_version":2}