{"id":"bda91157-0fb4-4bc5-aa7c-9d19ef3ac64a","arxiv_id":"2507.11265","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper proves explosion and nonexplosion criteria for exponential-transform Hawkes processes and establishes stability for nonpositive memory functions.","lead":"This paper analyzes Hawkes processes where the event rate is exp(ν + past influence), a model common in neuroscience. It gives conditions for when such processes blow up with infinitely many events in finite time, and when they settle into a stable, stationary pattern.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The explosion proof in Theorem 4.1(ii) drops the initial-condition term in equation (20); condition (14) is only pointwise, so the uniform lower bound used to obtain the infinite product (21) does not follow.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the explosion proof in Theorem 4.1(ii) silently omits the initial-condition contribution in equation (20), and condition (14) supplies only pointwise finiteness, not the uniform lower bound the product argument requires. The explicit construction above confirms that (14) can hold while A(t) is unbounded below near zero, so the gap is real rather than cosmetic. The nonexplosion part (i) appears internally coherent because it retains the initial-condition term and uses condition (13) to bound it. The stability part depends on an external theorem from Brémaud and Massoulié that is not independently verified in the preprint, but that is secondary to the explosion proof. Thus the reader's conditional verdict remains appropriate: the central explosion claim needs either an additional hypothesis or a repaired argument before the theorem is fully supported.","tokens_in":968,"tokens_out":992,"duration_ms":199469,"concrete_test":"Recompute equation (20) with the correct intensity lambda(t) = exp( nu + integral over (-inf,0] h(t-r) N0(dr) + sum_{j<k} h(t-T_j) ), and re-derive the product estimate (21) using only condition (14). Then plug in the explicit example: h(t) = 1 on (0, delta], h(t) = -n on [A_n + 1/(n+1), A_n + 1/n] with A_n = n^2, and N0 = sum_{n >= 1} delta_{-A_n}. Verify that (14) holds for every fixed t > 0 while A(t) = integral over (-inf,0] h(t-r) N0(dr) tends to -inf as t approaches 0 from above. If the factors in (21) are no longer bounded below by a positive constant, the explosion theorem is not proved by this argument; the stated hypotheses would need to be strengthened or the proof replaced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step in the explosion proof is equation (20). According to the construction (11), the conditional intensity on the history G_{k-1} is lambda(t) = exp( nu + integral over (-inf,0] h(t-r) N0(dr) + sum_{j=1}^{k-1} h(t-T_j) ), but (20) uses only the sum over previous self-exciting points and omits the initial-condition integral A(t) = integral over (-inf,0] h(t-r) N0(dr). The proof then bounds the self-exciting sum by (k-1) h_delta, which is valid on B_{k-1}, but no comparable lower bound is available for A(t). Condition (14) only asserts that A(t) > -inf for each fixed t > 0; it does not imply that the infimum of A(t) over t in (0, delta] is finite below. This is not a harmless technicality: take N0 = sum_{n >= 1} delta_{-A_n}, let h(t) = 1 for t in (0, delta], and let h(t) = -n on [A_n + 1/(n+1), A_n + 1/n] with A_n -> infinity. Then for each fixed t > 0 the negative contributions overlap for at most two values of n, so (14) holds, while A(t) -> -inf as t approaches 0 from above. Hence the lower bound exp(nu + (k-1)h_delta) in (20)-(21) is unjustified, and the infinite product may vanish. A repair requires either a uniform lower bound on A(t), a finite initial condition, or an entirely different argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces loglinear Hawkes processes, a nonlinear Hawkes model with conditional intensity lambda(t)=exp(nu + integral h(t-s) dN(s)), motivated by spike-response models in computational neuroscience. The authors construct the process via a Poisson-embedding/thinning argument, prove a uniqueness statement for the construction, and then establish sufficient conditions for nonexplosion and explosion in Theorem 4.1. For nonexplosion they assume a memory function whose values on (0,delta] are nonpositive and a bounded initial-condition contribution; for explosion they assume positive values on (0,delta] and pointwise finiteness of the initial-condition contribution. In Section 5 they prove stability in distribution for nonpositive memory functions with finite first absolute moment, using a reduction to Brémaud and Massoulié (1996), and derive necessary conditions for stability in Theorem 5.3. The paper closes with an outlook on extensions to partly positive memory functions and networks.","tokens_in":13411,"tokens_out":5741,"duration_ms":72948,"significance":"If the main theorems are correct, the paper fills a real gap: the exponential rate function is neither Lipschitz nor linearly dominated, so existing nonlinear-Hawkes stability and existence results do not apply. The construction is explicit and machine-checkable in style, with a clean Poisson-embedding proof and a clear uniqueness argument. The nonexplosion theorem is carefully argued, and the stability theorem for nonpositive h is a clean, direct application of a classical result with explicit hypotheses. The necessary condition (24) is a simple, falsifiable bound. The paper would be a useful theoretical foundation for loglinear Hawkes models in neuroscience and related fields. However, the proof of the explosion theorem contains a load-bearing gap, described below, that must be repaired before the result can be accepted as stated.","major_comments":[{"comment":"The proof of explosion drops the initial-condition term in the conditional probability calculation. In equation (20), the exponent contains only sum_{j=1}^{k-1} h(t-T_j), but the intensity in (11) is exp(nu + integral_{(-infinity,0]} h(t-s) N0(ds) + sum_{j=1}^{k-1} h(t-T_j)). The lower bound from (20) to (21) is therefore valid only if A(t)=integral_{(-infinity,0]} h(t-s) N0(ds) is uniformly bounded below on the interval (T_{k-1}, T_{k-1}+epsilon/k^2], which is contained in (0,delta] on the event B_{k-1}. Condition (14) only asserts pointwise finiteness for each fixed t>0; it does not imply a uniform lower bound. For instance, with N0=sum_n delta_{-A_n} and h(t)=1 on (0,delta] but h(t)=-n on [A_n+1/(n+1), A_n+1/n] for A_n tending to infinity, (14) holds for each fixed t>0, yet A(t) tends to -infinity as t approaches 0 from above. The infinite product in (21) can then be zero, so the claimed lower bound is unjustified. A repair requires either a uniform lower bound on A(t) over (0,delta], a finite initial condition, or an entirely different argument.","section":"Theorem 4.1(ii), equations (17)–(21)"}],"minor_comments":[{"comment":"There are small typos: 'choise' should be 'choice' and 'n (7)' should be 'in (7)'.","section":"Section 2"},{"comment":"The same symbol N is used for the constructed point process, the initial-condition counting measure N0, and the auxiliary Poisson random measure in the embedding; a brief notation remark would improve readability.","section":"Lemma 3.4 and surrounding text"},{"comment":"The display involving the dominating function tilde h is garbled in the manuscript and should be reformatted.","section":"Section 4, Example (iv)"},{"comment":"The condition is written as an inequality with '-infinity' without specifying 'almost surely' or 'for every t>0'; adding 'a.s.' and making the quantifier explicit would remove ambiguity, especially because the proof requires a uniform-in-t version.","section":"Theorem 4.1(ii), condition (14)"}],"recommendation":"major_revision","confidential_remarks":"The core contribution is valuable and the nonexplosion and stability parts appear sound, but the explosion theorem's proof has a genuine gap that is not merely cosmetic: condition (14) is too weak for the argument as written. I would encourage the authors to either strengthen the hypothesis to a uniform lower bound on the initial-condition contribution and state the theorem under that assumption, or provide a different proof that does not require such a bound. Given the otherwise careful exposition, I expect the gap can be fixed within the scope of a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Take a look at this one. The genuinely new content: for Hawkes processes with rate exp(nu + integral h dN), they give the first nonexplosion and explosion criteria and a stability result for nonpositive h. Nonexplosion theorem (4.1(i)) is well argued: local negativity of h on (0,delta] plus a one-sided bound on the initial-condition integral controls each interval. The reduction of stability for h<=0 to Bremaud-Massoulie via a bounded Lipschitz truncation is sensible, and the necessary condition via Jensen is clean.\n\nThe soft spot is Theorem 4.1(ii), the explosion criterion. Equation (20) conditions on G_{k-1}, which includes N0 and T_1...T_{k-1}, but the exponential in the conditional probability sums only over j=1,...,k-1. The initial-condition integral A(t)=int_{(-inf,0]} h(t-s)N0(ds) is missing. The hypotheses only give A(t)>-inf pointwise in t, not uniformly over t in (0,delta]. The stress-test example is valid: negative spikes at lags around 1/n make A(t) -> -inf as t decreases to 0 while (14) still holds. With A(t) arbitrarily negative near zero, the lower bound exp(nu + (k-1)h_delta) used in (20)-(21) does not follow, and the infinite product can vanish. The theorem may still be true; I would not be surprised if a different argument works, but the proof as written does not establish it. Fixing it likely requires a stronger condition (finite N0, or a uniform lower bound on A(t)) or a substantially different proof.\n\nThe rest is in decent shape. The construction in Lemma 3.4 includes the initial condition through the sum over all previous points, so the issue is confined to the proof of (ii). I could not independently verify that the hypotheses of Bremaud-Massoulie Theorem 2 are exactly met, but nothing in the reduction looks off.\n\nWho is this for: people working on nonlinear Hawkes processes, computational neuroscience models, and point-process stability. The nonexplosion and stability-for-inhibition parts are citable once the explosion proof is sorted. I would not cite it in its present form. I would send it to a serious referee; the topic is important and most of the paper is solid, but the referee should focus on Theorem 4.1(ii). A major revision is warranted.","headline":"Useful first results for loglinear Hawkes nonexplosion and stability, but the explosion proof has a real gap where the initial condition drops out.","tokens_in":13946,"tokens_out":7868,"would_cite":false,"duration_ms":92151,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G55","60G10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that the sign of the memory function near zero decides whether a loglinear Hawkes process explodes, and that nonpositive memory yields stability in distribution with a unique stationary limit.","keywords":["loglinear Hawkes process","nonlinear Hawkes process","explosion criterion","nonexplosion","stability in distribution","inhibition","exponential rate function","point process"],"falsifier":"Choose a memory function $h$ that is positive on $(0,\\delta]$ but takes large negative values at long lags, and an initial condition $N_0$ with points far in the past, so that $\\int_{(-\\infty,0]}h(t-r)N_0(dr)$ is finite for every $t>0$ yet tends to $-\\infty$ as $t\\downarrow0$. Simulate the Poisson-embedding construction of Lemma 3.4 many times and estimate the explosion probability. Theorem 4.1(ii) predicts $P(T_\\infty<\\infty)>0$ for every such pair satisfying (14); if any such pair fails to explode, the missing uniform lower bound is essential to the theorem as stated.","tokens_in":12839,"feed_emoji":"⚡","tokens_out":7910,"duration_ms":83823,"temperature":0.7,"pith_summary":"Loglinear Hawkes processes are point processes whose conditional intensity is $\\lambda(t)=\\exp\\{\\nu+\\int_{(-\\infty,t)}h(t-s)N(ds)\\}$, the exponential version of a nonlinear Hawkes process. The paper aims to give the first existence, explosion, and stability criteria for this model, which is not covered by earlier Lipschitz-based results because the exponential function is neither Lipschitz nor linearly bounded. Its main theorem shows that the sign of the memory function $h$ near zero decides explosion: if $h\\leq 0$ on an initial interval the process is nonexplosive, while if $h>0$ on an initial interval the process is explosive. For nonpositive $h$ with finite first moment, the dynamics are stable in distribution and admit a unique stationary version. A sympathetic reader should care because this is the standard 'spike response' model in computational neuroscience, and the paper supplies the missing theoretical footing for its use and simulation.","feed_headline":"Exponential Hawkes blow up unless memory is inhibitory near zero","feed_subtitle":"New criteria decide when loglinear Hawkes processes explode and when they converge to a unique stationary law.","key_machinery":"The carrying object is the exponential transfer function $\\varphi(x)=e^{\\nu+x}$, which maps the integrated past $\\int_{(-\\infty,t)}h(t-s)N(ds)$ to a positive rate. The construction uses Poisson embedding: take a unit-rate Poisson random measure on $\\mathbb{R}_{\\ge0}\\times\\mathbb{R}_{\\ge0}$ and keep points whose mark $z$ lies below $\\lambda(t)$, so the process is built recursively as $T_n=\\inf\\{t>T_{n-1}: N(\\{t\\}\\times[0,e^{\\nu+\\sum_{k<n}h(t-T_k)}])>0\\}$. Nonexplosion is proved by bounding $\\log\\lambda(t)$ on each interval $((k-1)\\delta,k\\delta]$ by a finite random variable $M_k$, so the number of points per interval is dominated by a Poisson count. Explosion is proved by lower-bounding the conditional probability of a small gap by $1-\\exp\\{-\\frac{\\varepsilon}{k^2}e^{\\nu+(k-1)h_\\delta}\\}$, whose product over $k$ is strictly positive exactly when $h_\\delta:=\\inf_{t\\in(0,\\delta]}h(t)>0$.","core_discovery":"On the paper's own terms, the central discovery is Theorem 4.1. For a loglinear Hawkes process with intensity $\\lambda(t)=\\exp\\{\\nu+\\int_{(-\\infty,t)}h(t-s)N(ds)\\}$ and initial condition $N_0$ on $(-\\infty,0]$, the paper proves: (i) if $\\sup_{t>0}h(t)<\\infty$, $h(t)\\leq0$ for $t\\in(0,\\delta]$, and $\\sup_{t>0}\\int_{(-\\infty,0]}h(t-r)N_0(dr)<\\infty$ a.s., then the process is nonexplosive ($T_\\infty=\\infty$ a.s.); (ii) if $\\inf_{t\\in(0,\\delta]}h(t)>0$ and $\\int_{(-\\infty,0]}h(t-r)N_0(dr)>-\\infty$ for every $t>0$, then the process is explosive ($T_\\infty<\\infty$ with positive probability). Theorem 5.2 then shows that for nonpositive $h$ with $\\int_0^\\infty t|h(t)|dt<\\infty$ and a mild decay condition on $N_0$, the process is stable in distribution with a unique stationary version. Theorem 5.3 gives necessary stability conditions, including the universal bound $\\int_0^\\infty h(s)ds\\le e^{-(1+\\nu)}$.","pith_inferences":["The paper's explosion proof in Theorem 4.1(ii) appears to use, without stating it, a uniform lower bound on the initial-condition contribution $\\int_{(-\\infty,0]}h(t-r)N_0(dr)$ for $t\\in(0,\\delta]$; condition (14) gives only pointwise finiteness. A natural test is whether processes with memory functions that have negative spikes at large lags and initial points far back still explode, as the theor","The stability result suggests a symmetry with linear Hawkes theory: just as linear Hawkes processes require nonnegative memory for their rate to stay positive, loglinear Hawkes processes require nonpositive memory for stability because the exponential turns any positive excursion into explosive feedback.","The universal bound $\\int_0^\\infty h(s)ds\\le e^{-(1+\\nu)}$ could be used as a model-checking diagnostic: estimated memory functions from stable neural spike trains should satisfy this integral constraint, and violations would indicate that the fitted loglinear model cannot be stationary."],"forward_implications":["If the nonexplosion part is correct, loglinear Hawkes processes with inhibitory memory near zero are legitimate models on the infinite time horizon: almost surely only finitely many events occur in any bounded interval, so simulation and likelihood-based inference are well-defined.","If the explosion part is correct, any memory function that is bounded below by a positive constant on some initial interval produces infinitely many events in finite time with positive probability, so such kernels cannot appear in stationary models.","If the stability theorem is correct, for nonpositive $h$ with $\\int_0^\\infty t|h(t)|dt<\\infty$, the long-run behaviour is insensitive to the initial condition: every solution converges to the same stationary loglinear Hawkes process.","The necessary stability bound $\\int_0^\\infty h(s)ds\\le e^{-(1+\\nu)}$ gives a concrete, checkable restriction on any memory kernel that is to yield a stationary version.","The Poisson-embedding construction provides an explicit recursive simulation algorithm for loglinear Hawkes processes, including cases with infinite initial conditions satisfying the theorem's hypotheses."],"supporting_citations":[{"why":"Introduces the linear Hawkes process that the loglinear model extends to an exponential rate function.","marker":"Hawkes (1971a)"},{"why":"Develops the self-exciting point process framework that motivates the general Hawkes intensity form.","marker":"Hawkes (1971b)"},{"why":"Supplies the nonlinear Hawkes framework, the Poisson-embedding Lemma 3, and Theorem 2 on stability to which Theorem 5.2 appeals.","marker":"Brémaud and Massoulié (1996)"},{"why":"Supplies the equivalence of weak convergence and convergence of finite-dimensional distributions for point processes used in the stability definition.","marker":"Daley and Vere-Jones (2003/2009)"},{"why":"Provides the infinite-product convergence criterion used to conclude that the lower bound (21) is positive in the explosion proof.","marker":"Knopp (1951, Chapter 7, Theorem 4)"}],"fun_headline_variants":["Memory sign near zero decides Hawkes blow-up","Inhibitory memory prevents loglinear Hawkes blow-up","Excitatory start forces explosive Hawkes","Loglinear Hawkes: sign near zero sets explosion","Memory positive near zero guarantees Hawkes blow-up"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The explosion half of the main theorem depends on the unstated assumption that the initial-condition contribution to the log-intensity, $\\int_{(-\\infty,0]}h(t-r)N_0(dr)$, stays uniformly bounded below for all $t$ in the first window $(0,\\delta]$; the theorem's stated condition (14) only says this integral is finite at each individual time, which is weaker and does not by itself justify the uniform lower bound used in the proof.","fun_headline_variants_meta":{"raw":{"variants":["Memory sign near zero decides Hawkes blow-up","Inhibitory memory prevents loglinear Hawkes blow-up","Excitatory start forces explosive Hawkes","Loglinear Hawkes: sign near zero sets explosion","Memory positive near zero guarantees Hawkes blow-up"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001284,"raw_usage":{"total_tokens":5225,"prompt_tokens":900,"completion_tokens":4325,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":4252}},"tokens_in":516,"tokens_out":4325,"duration_ms":38057,"temperature":1.0,"reasoning_tokens":4252,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:15:42.848093+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a memory function $h$ that is positive on $(0,\\delta]$ but takes large negative values at long lags, and an initial condition $N_0$ with points far in the past, so that $\\int_{(-\\infty,0]}h(t-r)N_0(dr)$ is finite for every $t>0$ yet tends to $-\\infty$ as $t\\downarrow0$. Simulate the Poisson-embedding construction of Lemma 3.4 many times and estimate the explosion probability. Theorem 4.1(ii) predicts $P(T_\\infty<\\infty)>0$ for every such pair satisfying (14); if any such pair fails to explode, the missing uniform lower bound is essential to the theorem as stated.","supporting_citations":[],"review_version":1}