{"id":"86b8df34-cb60-493b-ae89-1d083684ebb3","arxiv_id":"1908.05523","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors introduce a continuous-time self-exciting multifractional process as the solution of a Volterra stochastic differential equation and prove existence, uniqueness, Hölder regularity, and an Euler-Maruyama convergence rate.","lead":"A new class of random processes, self-exciting multifractional processes, is defined and analyzed. The paper proves such processes exist, are unique, have controlled regularity, and can be simulated with a known error rate.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reader's Lemma-4 objection does not land; the real defects are Lemma 15's false gamma H2 and the h(t,X_t) typo in Theorem 5.","rationale":"The reader's weakest assumption is only half right. The Lemma 4/H2 objection is incorrect: for the plain SEM kernel, the x-Lipschitz bound is |r^a-r^b| = |log r| r^\\xi |a-b|, so the problematic factor is present on both sides and (3.2) can hold with a large T-dependent constant. I therefore do not count it as a load-bearing gap in Theorem 5 or Proposition 8. The conditional verdict is still appropriate, but for different reasons: the central theorem's displayed equation is a genuine as-stated error (h(t,X_t) rather than h(t,X_s)), and the SEM-Gamma extension contains a false H2 inequality; the exponential damping term contributes a non-vanishing difference at lag one, so Lemma 15 and the resulting existence claim are unproved as written. Both are repairable, which is why I do not move the verdict to REJECT: adding a bounded term to k2 gives a valid K0 kernel, and changing X_t to X_s in (3.5) aligns the statement with the proof. The Section 4.2 examples with unbounded |x| and h_*=0 are outside the assumed parameter range and should be flagged as illustrative only.","tokens_in":19455,"tokens_out":31992,"duration_ms":309411,"concrete_test":"Check Lemma 15's (5.3) at t-s=1 with h≡3/4, f(x)=1+|x|, x=1, y=2: the left side is (e^{-2}-e^{-3})^2>0 and the printed right side is 0, so H2 as stated is false. Then verify whether adding a bounded term to k2, e.g. k2=C_T((t-s)^{2h_*-1}|log(t-s)|^2+1), is in K0 and restores the proof; if it does, the gamma existence claim is salvageable but the paper must supply it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the central SEM claim, Lemma 4's H2 estimate is not the weak point: writing r=t-s and a,b in [h_*-1/2, h^*-1/2], the mean value theorem gives |r^a-r^b| = |log r| r^\\xi |a-b|, so as r->1 both sides of (3.2) vanish at the same rate and a sufficiently large C_T makes the estimate true. The reader's near-one counterexample is therefore not available. The genuine as-stated weaknesses are: (i) Theorem 5 prints h(t,X_t) in (3.5) instead of h(t,X_s); taken literally the integrand is not F_s-adapted and Zhang's Theorem 2 cannot be invoked. (ii) Lemma 15's SEM-Gamma estimate (5.3) is false. At t-s=1 the RHS of (5.3) is zero, but for, e.g., h≡3/4 and f(x)=1+|x| the integrand equals e^{-f(t,x)}-e^{-f(t,y)}, which does not vanish for x≠y; the proof drops the |f(t,x)-f(t,y)||t-s| term. This makes the SEM-Gamma existence theorem unproved as written, although a repaired k2 containing a bounded term would satisfy H2. The Section 4.2 examples with h(x)=1/(1+x^2) or 1/2-1/(2(1+x^2)) also have infimum 0, outside Definition 3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a Self-Exciting Multifractional (SEM) process X^h defined as the solution of the Volterra equation X_t = g(t) + ∫_0^t (t-s)^{h(t,X_s)-1/2} dB_s, where h is a time- and path-dependent Hurst function. The authors prove existence and uniqueness via Zhang's theorem (Theorem 5), establish p-th moment bounds, show α-Hölder continuity for α < h_* ∧ δ (Proposition 8), and derive a strong Euler-Maruyama convergence rate |Δt|^γ for any γ < 2h_* (Theorem 9). A damped extension, the SEM-Gamma process, is defined by inserting an exponential factor e^{-f(t,X_s)(t-s)} into the kernel, with analogous existence, regularity, and numerical-scheme results claimed in Section 5. The paper ends with simulations and a discussion of applications to self-exciting phenomena such as earthquakes and financial crashes.","tokens_in":19749,"tokens_out":12944,"duration_ms":120230,"significance":"If the technical issues are repaired, the SEM class is a natural continuous-time counterpart to the Sornette–Filimonov discrete self-excited multifractional model, and the paper gives a serious application of Zhang's stochastic Volterra framework rather than a heuristic construction. The proof strategy is non-circular: existence and moment bounds reduce to external results (Zhang [16], the fractional Gronwall inequality of Ye–Gao–Ding) and standard inequalities, and no data fitting or target-driven construction is involved. The explicit convergence rate 2h_* for the Euler–Maruyama scheme is a useful quantitative feature. However, as written, the SEM-Gamma existence theorem is not proved, and several theorem statements contain errors that must be corrected before the results can be accepted.","major_comments":[{"comment":"Equation (3.5) in Theorem 5 writes h(t, X^h_t) inside the integrand, and equation (5.4) writes h(t, X^{h,f}_t) and f(t, X^{h,f}_t). Taken literally, the integrand is not F_s-adapted, and Zhang's Theorem 2 cannot be invoked; the displayed definition is therefore not the process whose existence is proved. The proof, Proposition 8, and the original equations (1.2) and (5.1) use h(t,X_s) (and f(t,X_s)), so the intended statement is clear, but the displayed equations must be corrected before the results can be taken as stated.","section":"Theorem 5, Eq. (3.5) and Section 5, Eq. (5.4)"},{"comment":"The SEM-Gamma Lipschitz estimate (5.3) is false as printed: for t-s=1 the right-hand side is zero, whereas the left-hand side equals |e^{-f(t,x)}-e^{-f(t,y)}|^2, which is nonzero for, e.g., f(t,x)=1+|x| and x≠y. In the proof, the term |f(t,x)-f(t,y)||t-s| arising from the exponential factor is dropped. Consequently the H2 hypothesis for Zhang's theorem is not verified for the SEM-Gamma kernel, and the existence statement for (5.4) is not proved. A repair is available by taking k2(t,s)=C_T[(t-s)^{2h_*-1}|\\log(t-s)|^2+1], which lies in K0, but this corrected kernel is not given in the paper.","section":"Lemma 15, Eq. (5.3)"},{"comment":"The functions h(x)=1/2 - 1/(2(1+x^2)) and h(x)=1/(1+x^2) take values arbitrarily close to 0, and the first satisfies h(0)=0, so they do not satisfy Definition 3, which requires [h_*,h^*] ⊂ (0,1) with h_*>0. The simulations in Figures 4.2 and 4.3 therefore correspond to a regime not covered by Theorems 5, 8, and 9. Please either restrict the examples to functions with a positive lower bound (e.g., by adding a small ε) or explicitly state that these simulations are outside the theoretical framework.","section":"Section 4.2, Examples"}],"minor_comments":[{"comment":"The moment bound in Theorem 5 displays the exponent h_*-1/2 in the integral term; consistency with k1 in Lemma 4 and with Zhang's Theorem 2 requires the exponent 2h_*-1.","section":"Theorem 5"},{"comment":"In the case |t-s|≥1 the proof bounds the difference by C_T|x-y|^2, but the displayed estimate (3.2) contains an additional |\\log(t-s)|^2 factor that vanishes at t-s=1. The conclusion is true—near t-s=1 the left side also vanishes at the rate |\\log(t-s)|^2—but the written proof should state this mean-value argument explicitly.","section":"Lemma 4, proof of (3.2)"},{"comment":"The bound |\\log(t-s)|^2 ≤ C (t-s)^{-2δ} is used for all s<t, but it only holds for t-s<1. The contribution from t-s≥1 is harmless and can be handled by splitting the integral; please make this explicit.","section":"Theorem 9, proof of (4.5)"},{"comment":"There are several typographical and reference issues: equation (5.4) repeats the X_t/X_s typo; the text before Lemma 16 says 'Theorem(18)' instead of 'Theorem 18'; reference [13] lacks volume and page details; and the phrase 'h = 1 any time the sample path crossed the x-axis again' is grammatically unclear and should be rewritten.","section":"Various"}],"recommendation":"major_revision","confidential_remarks":"The central SEM existence result is likely correct and the proof strategy is sound modulo the typos in the theorem statements. The main obstacle is Section 5: Lemma 15's H2 estimate is false as stated and needs a corrected kernel with a bounded term. Once that is fixed and the examples are brought within the hypotheses, the paper should be reconsidered. There is no circularity concern; the results reduce to external theorems. The manuscript is within the scope of the journal, and the self-exciting multifractional construction is sufficiently novel."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the paper introduces a genuinely new object—a continuous-time process whose Hurst exponent depends on the past of the process—and the main theorems on existence, moment bounds, Hölder regularity, and Euler–Maruyama convergence are largely correct. The most frequently cited objection, that Lemma 4's estimate (3.2) fails near t−s=1, does not hold up. Writing r=t−s, the mean value theorem gives |r^a−r^b|=|log r| r^ξ |a−b|, so both sides of (3.2) go to zero as r→1. The kernel k(t,s)=(t−s)^{2h_*−1} satisfies Zhang's H2 up to constants, and the central SEM existence theorem is in good shape.\n\nThe actual soft spots are fewer but real.\n\nFirst, Theorem 5 prints h(t,X_t) inside the integral in (3.5). Literally, that integrand is not F_s-adapted and Zhang's Volterra theorem cannot be invoked. The proof and the surrounding definition use X_s, so this is a typo, but it sits in a theorem statement and will mislead anyone who reads the display without the proof.\n\nSecond, the moment-bound exponent in Theorem 5 is wrong as printed. Zhang's H1 gives k1 ~ (t−s)^{2h_*−1}, but the display has (t−s)^{h_*−1/2}. Another typo, same fix.\n\nThird, the gamma extension has a real mathematical flaw. Lemma 15's estimate (5.3) is false. At t−s=1, the printed RHS is zero because log(1)=0, while the LHS, for a nonzero dampening function, is |e^{−f(t,x)}−e^{−f(t,y)}|^2, which does not vanish for x≠y. The appendix proof drops the term |f(t,x)−f(t,y)|(t−s) when bounding the difference. The SEM-Gamma existence claim is therefore unproved as written. This is repairable—add a bounded term to k2, or modify the estimate—but it needs to be done explicitly.\n\nFourth, the Section 4.2 examples, like h(x)=1/(1+x^2), have infimum 0, so they do not satisfy Definition 3's requirement h_*>0. The simulations are illustrative but sit outside the stated hypotheses.\n\nThe references look appropriate: Sornette–Filimonov for the motivating discrete model, Zhang for the Volterra machinery, and the standard mBm literature. I see no citation padding.\n\nWho should read this: probabilists working on stochastic Volterra equations or Hawkes-type self-excitation. The EM rate result is a nice addition. Applied readers should wait until the gamma section is repaired and the typos cleaned up; the SEM part already gives a solid continuous-time counterpart to the Sornette–Filimonov model.\n\nFor peer review: send it out. A competent referee can distinguish the typos from the one real gap, and the authors clearly have the ability to fix both. This is refereeable, not a desk reject.","headline":"The SEM construction is new and the main existence/EM results are sound once you fix two typos; the real mathematical gap is the gamma extension's false Lipschitz estimate, not Lemma 4.","tokens_in":20278,"tokens_out":6798,"would_cite":true,"duration_ms":61503,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H20","60G22","60G17","60H35"],"pacs":[],"model":"deepseek-v4-flash","headline":"A stochastic Volterra equation with a state-dependent power kernel is shown to have a unique solution, the self-exciting multifractional process.","keywords":["self-exciting process","multifractional Brownian motion","stochastic Volterra equation","Hurst function","Hölder regularity","Euler-Maruyama scheme","strong convergence","self-exciting multifractional gamma process"],"falsifier":"Evaluate inequality (3.2) at $t-s=2$ with $h_*=0.1$, $h(t,x)=0.9$, $h(t,y)=0.1$, and $|x-y|=1$: the left side is $(2^{0.4}-2^{-0.4})^2\\approx 0.315$, while the right side is $2^{-0.8}(\\log 2)^2\\approx 0.276$. The inequality fails at this one point, so the kernel does not satisfy the claimed Lipschitz control and the existence proof's load-bearing estimate is false.","tokens_in":19250,"feed_emoji":"","tokens_out":14398,"duration_ms":128295,"temperature":0.7,"pith_summary":"The paper proposes a continuous-time self-exciting multifractional process, defined as the solution of a stochastic Volterra equation whose kernel exponent depends on the process's own past. It proves that this equation has a unique solution, that the solution has finite moments of every order, and that its sample paths are $\\alpha$-Hölder continuous for every $\\alpha$ below the lower bound of the Hurst function. It then proves that the Euler--Maruyama discretization converges strongly at rate $|\\Delta t|^\\gamma$ for any $\\gamma < 2h_*$, so the process can be simulated with a controlled error. The aim is to give a rigorous continuous-time counterpart to the discrete self-excited multifractional model proposed for earthquakes and financial crashes, and to provide a family of processes whose local roughness changes with its own history.","feed_headline":"Proves a self-exciting roughness process exists","feed_subtitle":"A Volterra equation yields a unique process whose roughness depends on its own past, plus a convergent simulation scheme.","key_machinery":"The load-bearing machinery is a general existence and uniqueness theorem for stochastic Volterra equations with singular kernels, applied to $\\sigma(t,s,x)=(t-s)^{h(t,x)-1/2}$. The theorem requires linear growth and Lipschitz conditions controlled by a Volterra kernel $k\\in K_0$; the paper verifies these with $k(t,s)=C_T(t-s)^{2h_*-1}$, after bounding the difference of two power functions by a logarithmic factor. A second bound, Lemma 7, controls the time-increment of the kernel by an integrable factor $\\lambda_\\gamma(t,t',s)$, and a standard moment-based continuity criterion converts this into Hölder paths. For the Euler--Maruyama error, the paper uses a fractional Gronwall inequality to turn a Volterra integral bound on the mean-square error into the strong rate $|\\Delta t|^\\gamma$.","core_discovery":"The central claim is that the equation $X_t^h = g(t) + \\int_0^t (t-s)^{h(t,X_s^h)-1/2}\\,dB_s$, with $h$ a bounded Lipschitz function taking values in $(0,1)$, has a unique solution with finite moments of all orders, and this solution has $\\alpha$-Hölder paths for every $\\alpha < h_* \\wedge \\delta$ when $g$ satisfies the paper's continuity condition H4. The paper calls this solution a self-exciting multifractional process (SEM). The printed statement of Theorem 5 writes $h(t,X_t^h)$ in the integrand, but the definition in (1.2), the proof, and all subsequent estimates use $h(t,X_s^h)$; the latter is the operative equation. The same existence and regularity program is carried out for a damped version, the SEM-Gamma process, whose kernel is multiplied by $\\exp(-f(t,X_s^h)(t-s))$.","pith_inferences":["The state-Lipschitz estimate in Lemma 4 could be repaired by adding a bounded term that covers lags away from zero; the paper does not provide that correction, but the structure of the proof suggests the convergence rates would not change.","The rate dependence on $h_*$ suggests a practical diagnostic: estimating $h_*$ from data would directly quantify how finely a SEM model must be sampled for a given error tolerance.","The reported autocorrelation of absolute increments for SEM-Gamma comes from a single simulation; a systematic Monte Carlo study could test whether this volatility-clustering signal persists across seeds and parameter values."],"forward_implications":["The discrete self-excited multifractional model now has a rigorously defined continuous-time target, so simulation studies can be run against a well-posed process.","Because the SEM process has finite moments and Hölder paths with exponent controlled by $h_*$, it can serve as a noise model where the local roughness depends on the recent history of the process.","The Euler--Maruyama rate $\\gamma<2h_*$ gives a concrete rule: when the self-exciting dynamics produce a low lower-Hurst bound, smaller time steps are needed to keep the mean-square error under control.","The SEM-Gamma version inherits the same existence and convergence results, and its exponential damping produces a mean-reverting process with intermittency, making it a candidate for turbulence or volatility modeling."],"supporting_citations":[{"why":"supplies the Volterra existence and uniqueness theorem used to prove Theorem 5.","marker":"[16]"},{"why":"provides an Euler scheme for singular Volterra equations whose rate the paper sharpens.","marker":"[15]"},{"why":"is the discrete self-excited multifractional model that the SEM process is meant to make rigorous.","marker":"[13]"},{"why":"introduces the Gamma kernel construction that the SEM-Gamma extension builds on.","marker":"[1]"},{"why":"gives the fractional Gronwall inequality used to bound the Euler--Maruyama error.","marker":"[14]"},{"why":"is the source of the moment-based continuity criterion used for the Hölder path regularity.","marker":"[10]"}],"fun_headline_variants":["Self-exciting roughness: a new stochastic process","Volterra equation yields self-exciting multifractional paths","New process with past-dependent roughness proven unique","Euler-Maruyama converges for self-exciting rough process","Self-exciting multifractional process: existence and simulation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the kernel's dependence on the current state is controlled by the printed estimate for every time gap; that estimate is too small when the time gap exceeds one, so the existence proof as written depends on an inequality that fails.","fun_headline_variants_meta":{"raw":{"variants":["Self-exciting roughness: a new stochastic process","Volterra equation yields self-exciting multifractional paths","New process with past-dependent roughness proven unique","Euler-Maruyama converges for self-exciting rough process","Self-exciting multifractional process: existence and simulation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1508,"prompt_tokens":881,"completion_tokens":627,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":550}},"tokens_in":497,"tokens_out":627,"duration_ms":5430,"temperature":1.0,"reasoning_tokens":550,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:12:08.264045+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate inequality (3.2) at $t-s=2$ with $h_*=0.1$, $h(t,x)=0.9$, $h(t,y)=0.1$, and $|x-y|=1$: the left side is $(2^{0.4}-2^{-0.4})^2\\approx 0.315$, while the right side is $2^{-0.8}(\\log 2)^2\\approx 0.276$. The inequality fails at this one point, so the kernel does not satisfy the claimed Lipschitz control and the existence proof's load-bearing estimate is false.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Volterra existence and uniqueness theorem used to prove Theorem 5."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides an Euler scheme for singular Volterra equations whose rate the paper sharpens."},{"cited_title":"and Filimonov, V","cited_arxiv_id":null,"evidence_quote":"is the discrete self-excited multifractional model that the SEM process is meant to make rigorous."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces the Gamma kernel construction that the SEM-Gamma extension builds on."},{"cited_title":"and Ding, Y","cited_arxiv_id":null,"evidence_quote":"gives the fractional Gronwall inequality used to bound the Euler--Maruyama error."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the source of the moment-based continuity criterion used for the Hölder path regularity."}],"review_version":1}