{"id":"f5e16425-e952-438c-bbdd-f3b40284659f","arxiv_id":"1908.05906","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For Lévy risk processes with two-sided jumps, the optimal dividend and capital injection strategy is a double barrier rule: pay out the excess above a threshold and inject capital to prevent ruin.","lead":"This paper proves that a double barrier rule, pay out reserves above a chosen level and inject capital whenever reserves hit zero, is optimal for an insurance risk model whose reserves follow a Lévy process with both upward and downward jumps. It is the first rigorous optimality proof for this broad class and it extends known results that only covered jumps in one direction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unproved Assumption 2.1 leaves the optimality theorem conditional: the derivative formulas in Lemmas 5.3 and 5.6 require density regularity that Section 6 verifies only for unbounded variation processes with finite jump intensity on at least one side.","rationale":"The reader identified Assumption 2.1 as the weakest assumption, and my reading agrees. The proof of Theorem 5.1 is coherent within its stated assumptions, and I found no internal contradiction that would force a rejection. The most load-bearing point is indeed the unproved density regularity for unbounded variation processes with two-sided jumps: it is needed pointwise in Lemma 5.3 and for C^2 regularity in Lemma 5.6, but Section 6 only verifies the assumption when Π(-∞,0)<∞ or Π(0,∞)<∞. The abstract overstates the scope by omitting this condition. I do not see a decisive counterexample, so the appropriate verdict remains conditional: the theorem is plausible and well-structured, but the advertised generalization to two-sided jumps is not fully established. The self-referential sentence in Lemma 5.8 is a typographical-level issue, not a substantive circularity, because the surrounding argument only requires Lemma 5.7 to identify the region where (L-q)v vanishes. Thus my read does not change the reader's CONDITIONAL verdict.","tokens_in":25942,"tokens_out":12843,"duration_ms":132179,"concrete_test":"Choose a concrete unbounded variation Lévy process with infinite jump activity on both sides not covered by Section 6, e.g., X_t = B_t + J_t with J a symmetric tempered α-stable process, α∈(1,2), and compute φ_{a,0}(x) by solving the boundary value problem (L-q)u=0 on (0,a), u(0)=0, u(a)=1, using the known potential density of the killed process. If the solution has a locally bounded derivative that is continuous a.e., Assumption 2.1 holds for this example, showing the gap is a missing proof; if the derivative blows up or fails to exist on a set of positive measure, the theorem's stated scope is false and must be restricted. A definitive settlement would be an analytic proof or counterexample for the full class; the single-process computation decides whether the advertised two-sided infinite-activity regime is even plausible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 5.1 claims optimality of the double barrier strategy π_{a*} for Lévy processes satisfying Assumption 2.1. For unbounded variation paths, Assumption 2.1 postulates that the hitting-probability maps φ_{a,0}(x)=E_x[e^{-qτ_a^+};τ_a^+<τ_0^-] and φ_{0,a}(x)=E_x[e^{-qτ_0^-};τ_0^-<τ_a^+] have Radon–Nikodym densities that are continuous a.e. and locally bounded on (0,a). This is not a harmless technical condition: Lemma 5.3 computes the pointwise derivative v'_{π_a}(x)=φ_{a,0}(x)+βφ_{0,a}(x); Lemma 5.6 uses φ'_{a,0}, φ'_{0,a} to upgrade v_{π_{a*}} to C^2_line; Lemma 5.7 uses the continuity of Lv_{π_{a*}}. If Assumption 2.1 fails, these steps have no footing. The paper does not prove Assumption 2.1 for general two-sided jump processes: Section 6 verifies it only under Π(-∞,0)<∞ or Π(0,∞)<∞, i.e., finite jump intensity on at least one side. Thus the central claim, as advertised in the abstract for Lévy processes 'that may have positive and negative jumps,' is not established for processes with infinite jump activity on both sides. The assumption is stated explicitly, so Theorem 5.1 is internally valid conditional on it, but the scope of that condition is unknown. This is a correctness risk, not merely a presentation issue: if some unbounded variation process has a non-locally-bounded or discontinuous density, the verification argument in Section 5 does not apply. The paper gives no example of a two-sided infinite-activity process satisfying Assumption 2.1, and no proof that such processes exist beyond the one-sided finite-intensity cases. The self-referential 'By Lemma 5.8' inside the proof of Lemma 5.8 is almost certainly a typo for Lemma 5.7, since the displayed step only needs (L-q)v=0 on (0,a*); it is secondary to the Assumption 2.1 gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the classical de Finetti optimal dividend problem with capital injection for a Lévy risk process that may have two-sided jumps. The author proposes a double barrier strategy at a level a* defined by beta E_a[e^{-q kappa_{a,0}^-}] <= 1, where kappa is the first passage time of the reflected process below zero, and claims that this strategy is optimal among all admissible strategies. The proof strategy is to compute the derivatives of the expected net present value with respect to the initial state and the barrier level by a pathwise sample-path comparison, select a candidate barrier via a first-order condition, and then apply a verification lemma. Section 6 provides examples of unbounded variation Lévy processes satisfying the required regularity assumption, but only under a one-sided finite jump intensity condition. The main theorem is therefore conditional on a density regularity assumption that is not verified for general two-sided infinite-activity processes.","tokens_in":26352,"tokens_out":12795,"duration_ms":125310,"significance":"If the main theorem were established unconditionally, this would be the first proof of optimality of double barrier strategies for general two-sided jump Lévy processes, extending the spectrally negative and spectrally positive results of Avram et al. (2007) and Bayraktar et al. (2013). The proof technique is original: rather than scale functions, the paper uses pathwise comparisons of reflected processes to identify derivatives of value functions, and it employs a randomized hitting time to handle the boundary case where beta nu(a*) < 1. The paper is not machine-checked, but the main derivation is coherent and no circular fitting is apparent. The significance is substantial, but it is reduced by the fact that the central claim is conditional on an assumption whose scope is essentially open except for one-sided finite-activity cases.","major_comments":[{"comment":"Assumption 2.1 postulates that for unbounded variation Lévy processes the hitting-probability maps phi_{a,0} and phi_{0,a} admit Radon-Nikodym densities that are continuous almost everywhere and locally bounded on (0,a). This assumption is load-bearing: Lemma 5.3 computes v'_{pi_a} from these densities, Lemma 5.6 uses their derivatives to obtain C^2 regularity, and Lemma 5.7 relies on the continuity of Lv_{pi_{a*}}. Section 6 verifies Assumption 2.1 only when Pi(-infinity,0) < infinity or Pi(0,infinity) < infinity, i.e., when there is finite jump activity on at least one side. The abstract claims optimality for Lévy processes 'that may have positive and negative jumps' without this restriction, and the introduction advertises a general two-sided jump class. As written, Theorem 5.1 does not cover unbounded variation processes with infinite jump activity on both sides, and no argument is given that the density assumption holds there. This is a correctness-risk concern: if the density condition fails for some such process, the verification proof has no footing. The authors should either prove Assumption 2.1 in the general case or explicitly restrict the main theorem and abstract to the class where the assumption is known to hold.","section":"Assumption 2.1, Theorem 5.1, Section 6"},{"comment":"Lemma 5.7 begins with 'Suppose a* > 0', but Theorem 5.1 is stated for all a* defined by the infimum, and a* = 0 is possible for bounded variation processes (e.g., if beta nu(0+) <= 1). No separate proof is given for the case a* = 0. The verification lemmas, including the smoothness results in Lemma 5.6 and the inequality in Lemma 5.8, are developed under a* > 0, and the proof of Theorem 5.1 does not address the boundary case. If a* = 0, the strategy pi_0 is admissible only for bounded variation paths, and its value function is piecewise linear; the paper should either show that a* > 0 under Assumption 2.1 or supply a direct verification argument for a* = 0.","section":"Lemma 5.7 and Theorem 5.1"},{"comment":"In the proof of Lemma 5.8, after applying the Meyer-Ito formula and taking expectations, the text states 'By Lemma 5.8, we have' and then displays the key identity leading to (5.35). Since Lemma 5.8 is the very assertion being proved, this is a circular reference if read literally. It is likely a typo for Lemma 5.7, which justifies restricting the integral to [a* - epsilon, infinity), but as written the proof is self-referential and must be corrected.","section":"Proof of Lemma 5.8"}],"minor_comments":[{"comment":"In the proof of Lemma 3.2, the chain of inequalities bounding E|inf_{t in [0,u]} sum J_i| introduces a discounted Poisson integral with e^{-qt} and integrates over [0,infinity). Since the sum in question is not discounted and the infimum is over t <= u, the equality with the discounted integral appears to have the wrong sign or a missing factor; the undiscounted bound would be finite and sufficient. Please recheck the displayed inequality.","section":"Lemma 3.2, equation (3.6)"},{"comment":"Equation (5.7) contains a typo: the last term should be (R^{(x)}_t - R^{(x+epsilon)}_t), not (R^{(x)}_t - R^{(x)}_t).","section":"Lemma 5.3, equation (5.7)"},{"comment":"The randomized stopping times K^{p*}_0 and T^{p*}_0 are constructed using i.i.d. random variables A^{[n]}_p, but the enlarged probability space and the independence of these variables from the Lévy process X are not stated. Please make the probabilistic setup explicit.","section":"Lemma 5.5, step i)"},{"comment":"The statement 'It is easy to check that nu and nu_x are right continuous' is used to pass from the inequalities (4.7) and (4.9) to the limit in (4.1). A brief justification of this right-continuity would improve readability.","section":"Section 4, proof of Lemma 4.2"},{"comment":"The abstract and introduction claim optimality for Lévy processes with two-sided jumps without mentioning the extra density assumption for unbounded variation processes. Since the assumption is not verified in general, the claims should be qualified to match the actual hypotheses of Theorem 5.1.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper is a serious attempt at a longstanding open problem, and the proof method is original. The main obstacle is the unverified Assumption 2.1 for general two-sided infinite-activity processes. This is likely fixable either by proving the regularity using known potential-theoretic results or by narrowing the advertised class; it is also possible that the assumption fails in some cases, which would invalidate the claimed scope. The a*=0 boundary case and the apparent self-citation in Lemma 5.8 should also be addressed. I recommend major revision rather than rejection because the core derivation appears sound and the gaps are localizable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kei Noba has a serious result here, but the paper's own assumption takes some of the shine off the abstract. The core contribution is the first proof, among the cited literature, of optimality of a double barrier strategy for Lévy processes with genuine two-sided jumps, and the proof method is the real news: instead of scale functions, which do not exist in the two-sided setting, the author compares sample paths of reflected processes under shifted initial values and barriers, and reads off derivatives of the value functions from hitting-time Laplace transforms. That is a clever and transferable idea, and it is executed carefully across the appendices. The candidate barrier a* = inf{a>0: β E_a[e^{-qκ_{a,0}^-}] ≤ 1} is chosen by a first-order condition, and the verification scheme is coherent.\n\nThe soft spot is exactly the one the reader flagged, and it is not minor. Assumption 2.1 postulates for unbounded variation processes that certain hitting-probability maps have densities that are continuous a.e. and locally bounded. Lemma 5.3 and Lemma 5.6 lean directly on that, and Section 6 proves the condition only under Π(-∞,0)<∞ or Π(0,∞)<∞. So the theorem as stated covers general two-sided jump processes only conditionally on an unproved regularity property. The abstract omits the assumption entirely, which overstates the scope. If that density condition fails for some infinite-activity two-sided process, the verification argument does not apply. This is a correctness risk, not a presentation nit. The author should either prove the condition, narrow the theorem statement, or state it as a conjecture highlighted prominently.\n\nThere are smaller issues: the proof of Lemma 5.8 contains a self-referential \"By Lemma 5.8\" sentence, almost certainly a typo for Lemma 5.7; and Remark 2.3 asserts continuity of the φ maps without proof, then says the fact isn't important, which is slightly odd but not damaging. The citation pattern looks honest; the paper points at flaws in Yuen–Yin and Yin et al. and does not lean on self-citations. The main argument is internally coherent as far as I can tell, and no circularity or fitting is present.\n\nWho is this for: anyone working on de Finetti dividend problems, Lévy fluctuation theory, or singular control. It deserves a serious referee: the technique is novel and the claimed generalization fills a recognized gap. My recommendation is to engage with it after the Assumption 2.1 gap is addressed; the paper should not be taken as proven for general two-sided Lévy processes as the abstract suggests.","headline":"A genuinely new pathwise comparison method proves optimality of double barrier strategies for two-sided jump Lévy processes, but the main theorem is conditional on an unproved density assumption that the abstract fails to advertise.","tokens_in":26932,"tokens_out":1740,"would_cite":true,"duration_ms":16942,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G51","93E20","91G80"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that the double barrier strategy at the level $a^*$ is optimal for the optimal dividend problem with capital injection when the risk process is a Lévy process with two-sided jumps, under Assumption 2.1.","keywords":["optimal dividend problem","capital injection","double barrier strategy","Lévy process with two-sided jumps","verification lemma","reflected Lévy process","expected net present value","hitting-time Laplace transform"],"falsifier":"Find an unbounded-variation Lévy process with infinite jump activity on both sides and compute, for some $a>0$, the density of $x\\mapsto E_x[e^{-q\\tau_a^+};\\tau_a^+<\\tau_0^-]$ or $x\\mapsto E_x[e^{-q\\tau_0^-};\\tau_0^-<\\tau_a^+]$; if that density is discontinuous or unbounded on $(0,a)$, Assumption 2.1 fails and the verification proof in Section 5 does not apply to that process.","tokens_in":25690,"feed_emoji":"💰","tokens_out":10572,"duration_ms":93594,"temperature":0.7,"pith_summary":"This paper establishes that, for a Lévy risk process whose jumps may go both up and down, the optimal dividend-and-bailout policy is a double barrier strategy: pay out any surplus above a fixed level $a^*$, and inject capital whenever the surplus would fall below zero. The optimal level is $a^*=\\inf\\{a>0:\\beta E_a[e^{-q\\kappa_{a,0}^-}]\\le 1\\}$, where $\\beta>1$ is the unit cost of injected capital and $\\kappa_{a,0}^-$ is the first time a doubly reflected process starting at $a$ hits zero. If the proof is right, no other admissible strategy achieves a higher expected value, defined as expected discounted dividends minus $\\beta$ times expected discounted injections. The result carries the known optimality for spectrally one-sided Lévy processes over to genuinely two-sided jumps, using pathwise comparisons instead of scale functions.","feed_headline":"Double barrier rule is optimal for two-sided-jump Lévy risk models","feed_subtitle":"One payout-and-bailout threshold maximizes expected dividends minus injection costs, even with jumps both ways.","key_machinery":"The key object is the value $v_{\\pi_a}$ of a double barrier strategy together with its one-sided derivatives with respect to the initial surplus and the barrier level. Because two-sided jumps destroy scale-function representations, the paper obtains these derivatives by coupling the reflected processes with initial value $x$ and $x+\\epsilon$ (or barrier $a$ and $a+\\epsilon$) on the same driving path and tracking how dividends and injections change; the changes are at most $\\epsilon$, so the derivatives fall out as Laplace transforms of first hitting times. Concretely, $v'_{\\pi_a}(x)=\\varphi_{a,0}(x)+\\beta\\varphi_{0,a}(x)$ on $(0,a)$, where $\\varphi_{a,0}(x)=E_x[e^{-q\\tau_a^+};\\tau_a^+<\\tau_0^-]$ and $\\varphi_{0,a}(x)=E_x[e^{-q\\tau_0^-};\\tau_0^-<\\tau_a^+]$, and $\\partial_a v_{\\pi_a}(x)=-\\nu_x(a)(1-\\beta\\nu(a))/(1-\\nu(a)\\nu_0(a))$. These formulas locate the maximizing barrier and give the concavity and derivative bounds needed by the verification lemma.","core_discovery":"The central claim is Theorem 5.1: under Assumption 2.1, the double barrier strategy at $a^*$ is optimal, and the value function of the problem is $v=v_{\\pi_{a^*}}$. The paper identifies $a^*$ through the Laplace transform of the first passage time of the reflected process, then proves the candidate satisfies a verification lemma: $v_{\\pi_{a^*}}$ has derivative between $1$ and $\\beta$, belongs to the required smoothness class, and satisfies the generator inequality $\\mathcal{L}w-qw\\le 0$ on $(0,\\infty)$. This supplies the first optimality proof for a general strategy in this dividend/capital-injection problem when the Lévy process has two-sided jumps; earlier claimed proofs for general two-sided-jump processes had gaps.","pith_inferences":["An extension not pursued here: for finite-horizon or random-horizon versions of the problem, the constant barrier $a^*$ would likely become a time-dependent boundary, and the same pathwise-shift method could yield the needed derivative identities for a verification argument.","The same pathwise-shift technique may prove optimality in other two-sided singular control problems, such as versions with transaction costs or with different cost rates for injections, whenever the hitting-time Laplace transforms have the required density regularity.","A numerical test of Assumption 2.1 is available: for an unbounded-variation process with infinite jump activity on both sides, such as a tempered stable process, one can estimate the densities of $\\varphi_{a,0}$ and $\\varphi_{0,a}$ by Monte Carlo simulation of small shifts; a violation would appear as unstable or unbounded derivative estimates near some interior point."],"forward_implications":["For every Lévy process satisfying Assumption 2.1, the optimal value in the dividend/capital-injection problem is exactly $v_{\\pi_{a^*}}$, so the search over all admissible strategies reduces to choosing one number $a^*$.","The optimal barrier is the smallest $a$ for which $\\beta E_a[e^{-q\\kappa_{a,0}^-}]\\le 1$; equivalently, raise the dividend barrier until the expected discounted cost of a future bailout, weighted by $\\beta$, no longer offsets the extra dividend income.","The verification lemma provides a practical certificate: any candidate function $w$ with $w'\\in[1,\\beta]$ and $\\mathcal{L}w-qw\\le0$ on $(0,\\infty)$ dominates all admissible strategies.","Together with the spectrally one-sided cases, the result covers Lévy risk processes with bounded variation, mixed-exponential jump diffusions, and unbounded-variation processes with one side of the jump measure finite."],"supporting_citations":[{"why":"Supplies the double barrier construction and the verification-lemma framework that the paper adapts.","marker":"[2]"},{"why":"The spectrally positive case whose result is generalized, providing the dual-model counterpart.","marker":"[4]"},{"why":"Used in Lemma 5.7 for the martingale argument identifying the generator applied to the candidate value.","marker":"[5]"},{"why":"Provides scale-function properties used in Section 6 to verify Assumption 2.1 for unbounded-variation processes with one finite side of jumps.","marker":"[8]"},{"why":"Supplies standard Lévy process facts, including the integrability condition behind Assumption 2.1 and reflection notation.","marker":"[9]"},{"why":"The semimartingale change-of-variable formula used in the verification lemma and in Lemma 5.8.","marker":"[16]"}],"fun_headline_variants":["Two-sided jump Lévy? Double barrier is optimal","Double barrier proven optimal for two-sided Lévy jumps","Optimal dividends with bailouts: double barrier for Lévy","First proof: double barrier optimal for two-sided Lévy","Lévy processes with jumps both ways: double barrier wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 2.1: when the Lévy process has unbounded variation paths, the hitting-time Laplace transforms $\\varphi_{a,0}$ and $\\varphi_{0,a}$ must have Radon–Nikodym densities that are continuous almost everywhere and locally bounded on $(0,a)$, and the paper verifies this only when at least one side of the jump measure is finite.","fun_headline_variants_meta":{"raw":{"variants":["Two-sided jump Lévy? Double barrier is optimal","Double barrier proven optimal for two-sided Lévy jumps","Optimal dividends with bailouts: double barrier for Lévy","First proof: double barrier optimal for two-sided Lévy","Lévy processes with jumps both ways: double barrier wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1286,"prompt_tokens":855,"completion_tokens":431,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":364}},"tokens_in":471,"tokens_out":431,"duration_ms":4379,"temperature":1.0,"reasoning_tokens":364,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:02:24.177245+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find an unbounded-variation Lévy process with infinite jump activity on both sides and compute, for some $a>0$, the density of $x\\mapsto E_x[e^{-q\\tau_a^+};\\tau_a^+<\\tau_0^-]$ or $x\\mapsto E_x[e^{-q\\tau_0^-};\\tau_0^-<\\tau_a^+]$; if that density is discontinuous or unbounded on $(0,a)$, Assumption 2.1 fails and the verification proof in Section 5 does not apply to that process.","supporting_citations":[{"cited_title":"Avram, Z","cited_arxiv_id":null,"evidence_quote":"Supplies the double barrier construction and the verification-lemma framework that the paper adapts."},{"cited_title":"Bayraktar, A","cited_arxiv_id":null,"evidence_quote":"The spectrally positive case whose result is generalized, providing the dual-model counterpart."},{"cited_title":"Biffis and A","cited_arxiv_id":null,"evidence_quote":"Used in Lemma 5.7 for the martingale argument identifying the generator applied to the candidate value."},{"cited_title":"Kuznetsov, A","cited_arxiv_id":null,"evidence_quote":"Provides scale-function properties used in Section 6 to verify Assumption 2.1 for unbounded-variation processes with one finite side of jumps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The semimartingale change-of-variable formula used in the verification lemma and in Lemma 5.8."}],"review_version":1}