{"id":"5cab9571-24ce-41cc-8d7d-8cd2eb5faffa","arxiv_id":"2506.15037","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For 2BSDEs with an exogenous random stopping time, the paper shows well-posedness and comparison under a shared conditional-density assumption, then links the equations to stochastic control and to a fully nonlinear PDE.","lead":"Backward equations that compute a value by working from a final payoff are extended to the case where the final time is a random 'default' event the equation cannot foresee. The paper proves existence, uniqueness, and comparison for these equations and applies them to control problems with ambiguous volatility and uncertain horizon.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The P0-density Hypothesis 2.5 is load-bearing, and the Girsanov example given for it is invalid: independence of Z_T/Z_t from F_t does not yield the displayed equality unless τ is independent of the density; only the independent-τ example is rigorously supported.","rationale":"The reader's conditional verdict is appropriate. The most load-bearing point is indeed Hypothesis 2.5, because without it the processes λ and U in the erratic-horizon 2BSDE (4) are not defined on the common filtration and the reduction in Theorem 1 has no footing. I add a concrete defect in the paper's evidence that the hypothesis is non-vacuous: Example (1) claims a Girsanov invariance of the conditional law of τ that requires more than Z_T/Z_t ⊥ F_t; it requires Z_T/Z_t to be independent of F_t jointly with τ, or τ independent of F. The closed-form test with τ=(W_1)^2 shows the claimed identity is false. This does not invalidate Theorem 1, which is conditional on the Hypothesis and follows established 2BSDE lines from [62, 64]; but it does mean the paper's stated applicability to correlated erratic horizons is not supported by the example. Other issues (the sign typo in Theorem 1's proof and the omitted proof of Theorem 4) are secondary and already noted by the reader. The verdict remains CONDITIONAL: the authors should correct or qualify Example (1), give a genuinely non-independent example if one exists, and fix the presentation issues.","tokens_in":27360,"tokens_out":12315,"duration_ms":118504,"concrete_test":"For T=1, take P0 as Wiener measure, F_t the Brownian filtration, τ=(W_1)^2, and define P^α by dP^α/dP^0 = exp(W_1−1/2), so Z_1/Z_0 is independent of F_0 (constant drift μ=1). Check Hypothesis 2.5 at t=0, x=1 for P^α: the left side is E^0[e^{W_1−1/2} 1_{W_1^2≥1}] = ∫_{|w|≥1} φ(w−1) dw = P(N(1,1)≤−1) + P(N(1,1)≥1) ≈ 0.0228 + 0.5 = 0.5228, while P^0(W_1^2≥1) = P(|W_1|≥1) ≈ 0.3173. These differ, so the equality asserted in Example (1) fails; hence that Girsanov family does not satisfy Hypothesis 2.5 unless τ is independent of the Brownian motion. Re-running this one closed-form check settles whether the paper's non-vacuity evidence for the density hypothesis is correct.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Hypothesis 2.5 is the single mechanism that turns the measure-dependent compensator of H and the measure-dependent jump size U into objects defined on the common filtration G^{P0+}; without it, the reduction of (4) to the auxiliary 2BSDE (6) in Theorem 1 cannot be formulated, and Remark 3.10's aggregation of U collapses. The paper's non-vacuity evidence is therefore critical. Example (1) claims that for P^α ∼ P^0 with dP^α/dP^0 = Z_T and Z_T/Z_t independent of F_t, one has P^α(τ≥x|F_t) = E^0[Z_T/Z_t 1_{τ≥x}|F_t] = P^0(τ≥x|F_t). The last equality is not valid: independence of Z_T/Z_t from F_t alone does not make the conditional expectation of (Z_T/Z_t)1_{τ≥x} equal to E^0[Z_T/Z_t] E^0[1_{τ≥x}|F_t]; one would need Z_T/Z_t independent of σ(F_t ∨ σ(τ)). For any τ that is a nontrivial functional of the Brownian path, the two conditional expectations differ, so the P0-density hypothesis fails for the displayed Girsanov family. Thus the only rigorously supported example is τ independent of F with an invariant law (Example 2), which is a narrow subfamily. This does not disprove Theorem 1, but it means the paper's stated evidence that the hypothesis admits 'two structurally different families' is incorrect, and the claimed applicability to erratic environments where the default is correlated with the market is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a second-order BSDE with an exogenous random horizon T∧τ, where τ is a random time enlarging the Brownian filtration. Under a P0-density hypothesis that a single conditional density kernel for τ works simultaneously for every measure in a non-dominated family P0, it states existence and uniqueness of a solution (Theorem 1 and Corollary 3.7), a comparison theorem (Theorem 2), and a piecewise Feynman-Kac connection to a fully nonlinear PDE in the Markovian case (Theorem 3). It then applies these results to erratic stochastic control problems with controlled volatility and with adversarial volatility (Theorems 4 and 5), and includes an appendix on CARA utility. The main strategy is to reduce the random-horizon 2BSDE to a fixed-horizon auxiliary 2BSDE and to give the explicit decomposition Y_t=Y^b_t 1_{t<τ}+ξ^a_τ 1_{t≥τ}, Z_t=Z^b_t 1_{t<τ}, U_t=(ξ^a_t-Y^b_t)1_{t<τ}.","tokens_in":2202,"tokens_out":2345,"duration_ms":251679,"significance":"If the auxiliary results from the cited preprint [62] are valid, the paper proposes a genuine extension of 2BSDE theory to exogenous random horizons, and the explicit decomposition (7) is a clean and potentially useful structural result. The aggregation of the jump-size process U across a non-dominated family, highlighted in Remark 3.10, is also a noteworthy contribution. However, the central existence engine is imported from an unreviewed preprint, one of the two advertised examples of the key hypothesis is incorrect, a main proof is omitted, and there are sign inconsistencies in the central verification. The theoretical framework is original but the manuscript, in its present form, is not sufficiently self-contained or rigorous for publication.","major_comments":[{"comment":"The displayed verification of the primal equation (4) on the events {τ>T} and {τ∈(t,T]} uses the wrong sign for the driver integral. Subtracting equation (6) at time τ from equation (6) at time t gives Y^b_t = Y^b_τ − ∫_t^τ F^{P,b}_s(...)ds − (∫_t^τ Z^b_s·dX^{c,P}_s)^P − (M^{b,P}_τ − M^{b,P}_t) + (K^{b,P}_τ − K^{b,P}_t), but the proof writes a plus sign before the F-integral. The subsequent replacement of Y^b_τ by ξ^a_τ and the intermediate term −(ξ^a_τ−Y_τ) do not follow from the displayed equalities. Because (4) also carries a minus sign before F, the computation as written does not establish that (7) solves (4). This is a load-bearing point in the proof of Theorem 1 and must be corrected.","section":"Section 3.1, Theorem 1, Step 2"},{"comment":"Lemma 3.1 is the central existence result for the auxiliary 2BSDE (6), upon which Theorem 1 rests, but its proof consists of verifying a few conditions and then citing the unreviewed preprint [62, Theorem 3.6 and Corollary 3.7]. The verification of the hypotheses of the cited theorem is not complete: the text says that Assumption 1 directly gives [62, Assumption 2.20 (ii)] and defines r^2_t:=C(1+λ_t)^2 for (iv), but does not check the remaining structural conditions such as the concatenation and universal measurability properties required by [62]. Please give a full proof of the auxiliary existence result, or state the cited theorem completely and verify every hypothesis, or ensure that [62] is available in published form before this result is used.","section":"Section 3.1, Lemma 3.1"},{"comment":"The Girsanov example under Hypothesis 2.5 is invalid. The equality P^α(τ≥x|F_t)=E^0[Z^α_T/Z^α_t 1_{τ≥x}|F_t] is correct, but the next equality E^0[Z^α_T/Z^α_t 1_{τ≥x}|F_t]=P^0(τ≥x|F_t) does not follow from the stated assumption that Z^α_T/Z^α_t is independent of F_t. One would need conditional independence of the density ratio from τ given F_t, which is a much stronger requirement. If τ is a nontrivial functional of the Brownian path, the F_t-conditional law of τ changes under the Girsanov change of measure and the P0-density Hypothesis fails. Thus the paper's claim that the hypothesis admits 'two structurally different families' is not supported by the displayed example; the independent-τ example (2) is the only one rigorously established.","section":"Section 2.2, Example (1)"},{"comment":"The minimality condition in Definition 2.9(2) is only verified on the event A^+={τ>T}: the final displayed equality in Step 4 is stated for ω∈A^+. On the event {τ∈(t,T]}, which has positive probability in general, one has K^{P'}_{T∧τ}−K^{P'}_{t∧τ}=K^{b,P'}_τ−K^{b,P'}_t, and the argument given does not apply. Since the minimality condition is required P-a.s. for every deterministic t, this leaves a gap in the existence proof. Please supply an argument showing that the minimality of K^b passes to the stopped differences, for example through optional stopping or through an explicit use of the conditional density of τ given F_t.","section":"Section 3.1, Theorem 1, Step 4"},{"comment":"The proof of Theorem 4 is explicitly omitted ('We omit the proof regarding the optimizers...'), yet this is one of the two main applications of the paper. The value-function identity V0(ξ)=sup_α sup_{(P,β)∈N^α} E^P[Y_0] and the characterization of optimal controls are not derived in the text. The references to the proof of Theorem 5 and to [19, Proposition 5.4] are not a substitute, since the erratic-horizon setting changes the admissible set and the form of the 2BSDE. Please provide a complete proof or state Theorem 4 as a corollary with all missing steps supplied.","section":"Section 4.3, Theorem 4"},{"comment":"There is an unannounced sign change between the theory of Section 3 and the control applications. The existence theory of Section 3 is developed for equations of the form (4), where the driver appears with a minus sign and an orthogonal martingale term dM^P is present. The control 2BSDEs (16)-(17), in contrast, are written with a plus sign before the driver, without the M term. As written, Theorem 1 and Corollary 3.7 do not apply verbatim to (16)-(17). The paper should explicitly state the transformation F↦−F (and the reduction M=0) that brings (16)-(17) into the form (4), or prove the existence result directly for the sign convention used in the control section.","section":"Section 4, equations (16)-(17)"}],"minor_comments":[{"comment":"The first abstract states that the driver is 'Lipschitz continuous', while the full-text abstract says 'Lipschitz continuous in y,z and stochastic Lipschitz in the jump u'. Please harmonize the two statements.","section":"Abstract"},{"comment":"The reference measure P^0 and the family P0 share a similar notation, which makes statements such as 'P0 is composed by all the probability P^α' confusing. Please use a distinct notation for the reference measure.","section":"Section 2.2, Example (1)"},{"comment":"The brief reference to [62, Section 3.1] for universal measurability is terse. Since the aggregation of U in Remark 3.10 depends on measurability, one or two sentences explaining the measurability argument would improve readability.","section":"Section 3.1, Remark 3.2"},{"comment":"In part (i) of the proof, the existence statement 'follows the same lines as the proof of Theorem 1 above' is vague. For a fixed P this is standard in the BSDE-with-jumps literature, but a precise citation or a short outline of the existence argument would be helpful.","section":"Section 3.2, Lemma 3.5"},{"comment":"The assertion that 'Doob's maximal inequality together with (27) gives ∥Z^n_t∥ x\\mid F_t)$ are different and no single kernel $\\gamma$ represents both simultaneously. Then the intensity is measure-dependent, the jump-size process $U$ cannot be aggregated, and Theorem 1 has no object to attach to, showing that the binding restriction is the hypothesis rather than the equation itself.","tokens_in":27138,"feed_emoji":"⏳","tokens_out":8677,"duration_ms":80940,"temperature":0.7,"pith_summary":"This paper introduces and solves a class of second-order backward stochastic differential equations (2BSDEs) in which the terminal time is an exogenous random default time, possibly invisible to the Brownian filtration. The central result is existence, uniqueness, and comparison for such equations, provided a single conditional density kernel for the default time is shared by every probability measure in the non-dominated family. The solution takes an explicit before/after-default form: before default it follows a fixed-horizon auxiliary 2BSDE, and after default it freezes at a pre-specified post-default payoff. In the Markovian case the same solution is represented through a fully nonlinear PDE, with the random horizon absorbed into a piecewise diffusion-expectation representation. The framework is applied to erratic stochastic control, covering both an agent who controls drift and volatility and an agent facing worst-case volatility chosen by an adversarial Nature.","feed_headline":"Random-horizon 2BSDEs solved by freezing at default time","feed_subtitle":"A single shared default-density kernel tames the end-time jump and links erratic control to a fixed-horizon equation.","key_machinery":"The load-bearing object is the $P_0$-density hypothesis: a single $F$-adapted kernel $\\gamma(t,u)$ must represent the conditional survival probability $P(\\tau > x \\mid F_t) = \\int_x^\\infty \\gamma(t,u)\\,du$ simultaneously under every probability measure in the non-dominated family $P_0$. This makes the intensity $\\lambda_t = \\gamma(t,t)/P(\\tau>t\\mid F_t)$ and the jump-size process $U$ well-defined across all $P$, so the random-horizon 2BSDE can be reduced to a Brownian-driven auxiliary 2BSDE on the fixed interval $[0,T]$; the jump is then reintroduced through the before/after-default decomposition of the terminal data.","core_discovery":"The central claim is that, under a Lipschitz driver and a $P_0$-density hypothesis, the 2BSDE with erratic horizon (4) has a unique solution in the sense of Definition 2.9, and that solution has the explicit before/after-default form $Y_t = Y^b_t \\mathbf{1}_{t<\\tau} + \\xi^a_\\tau \\mathbf{1}_{t\\ge\\tau}$, $Z_t = Z^b_t \\mathbf{1}_{t<\\tau}$, $U_t = (\\xi^a_t - Y^b_t)\\mathbf{1}_{t<\\tau}$, where $(Y^b,Z^b,K^b)$ is the unique solution of the fixed-horizon auxiliary 2BSDE (6). The same construction yields a comparison principle: if terminal payoffs are ordered and post-default payoffs coincide, the corresponding solutions are ordered. In the Markovian case the solution is $Y_t = v(t,X_t)\\mathbf{1}_{t<\\tau} + g(\\tau,X_\\tau)\\mathbf{1}_{t\\ge\\tau}$, with $v$ solving a fully nonlinear PDE that does not itself contain the random horizon; the default time enters only through the piecewise representation and the boundary value.","pith_inferences":["The paper does not test this, but if the $P_0$-density hypothesis holds, numerical methods developed for fixed-horizon 2BSDEs transfer directly to erratic-horizon problems: solve the auxiliary equation, then freeze at the sampled default time.","The structure suggests that uncertainty about timing and uncertainty about volatility separate cleanly, one entering through the boundary piece and the other through the PDE Hamiltonian, so in applications the default kernel and the volatility ambiguity could be estimated separately.","Beyond the paper, the $P_0$-density hypothesis could be relaxed by allowing measure-dependent intensities if one is willing to keep a family of jump-size processes indexed by probability; the aggregation of $U$ across $P_0$ is the real novelty of the construction, not the jump itself.","A testable implication for finance or cyber-risk settings is that if sudden exits are modeled with one shared conditional density across models, robust optimal strategies are determined by the fixed-horizon solution up to default; comparing this against a model with measure-dependent default intensity would quantify the cost of the hypothesis."],"forward_implications":["Any Lipschitz erratic-horizon 2BSDE satisfying the $P_0$-density hypothesis can be solved by solving a fixed-horizon auxiliary 2BSDE on $[0,T]$ and freezing the value at the post-default payoff after $\\tau$.","The comparison principle turns the solution into an ordering-preserving nonlinear expectation under the non-dominated family, so monotone inputs yield monotone values.","In the Markovian setting the random horizon drops out of the PDE itself: the value function solves a fully nonlinear PDE independent of $\\tau$, with the default time entering only through the piecewise representation and the boundary term $g(\\tau,X_\\tau)$.","Both erratic control problems, full volatility control and adversarial Nature ambiguity, have their value functions represented by the initial value of the corresponding 2BSDE, and optimality is characterized by pointwise attainment of the Hamiltonian together with $K_{T\\wedge\\tau}=0$ under the optimal measure."],"supporting_citations":[{"why":"Supplies the before/after-default decomposition of erratic-horizon BSDEs that Theorem 1 extends to the second-order setting.","marker":"[41]"},{"why":"Gives the reduction of a random-horizon BSDE to a Brownian BSDE with a singular driver, used in the auxiliary construction.","marker":"[35]"},{"why":"Provides the existence and uniqueness result for the auxiliary 2BSDE on which Lemma 3.1 relies.","marker":"[62]"},{"why":"Provides the non-dominated 2BSDE framework and dynamic programming principle used throughout the paper.","marker":"[64]"},{"why":"Supplies the conditional-density approach and the survival formula $P(\\tau>t\\mid F_t)=e^{-\\Lambda_t}$ that motivates the $P_0$-density hypothesis.","marker":"[23]"},{"why":"Supplies the classical 2BSDE-to-PDE link used in Lemma 3.12 for the Markovian representation.","marker":"[69]"},{"why":"Supplies the weak-formulation volatility-control framework that the erratic stochastic control application extends.","marker":"[19]"},{"why":"Supplies the volatility-ambiguity and worst-case framework that the sup-inf application extends.","marker":"[27]"}],"fun_headline_variants":["Freezing at default time tames erratic-horizon 2BSDEs","Random-horizon 2BSDEs: comparison principle via default-time freezing","Erratic-horizon control: 2BSDE solution via single PDE","Default-time freezing solves 2BSDE with uncertain horizon","Stochastic control under vague deadlines tamed by 2BSDE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the $P_0$-density hypothesis: one conditional density kernel for the default time must be valid simultaneously under every probability measure in $P_0$; if the kernel depends on the measure, the intensity and jump process cannot be aggregated and the reduction to the auxiliary 2BSDE collapses.","fun_headline_variants_meta":{"raw":{"variants":["Freezing at default time tames erratic-horizon 2BSDEs","Random-horizon 2BSDEs: comparison principle via default-time freezing","Erratic-horizon control: 2BSDE solution via single PDE","Default-time freezing solves 2BSDE with uncertain horizon","Stochastic control under vague deadlines tamed by 2BSDE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000694,"raw_usage":{"total_tokens":3161,"prompt_tokens":992,"completion_tokens":2169,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":2072}},"tokens_in":608,"tokens_out":2169,"duration_ms":14507,"temperature":1.0,"reasoning_tokens":2072,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:45:28.498602+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find a non-dominated family $P_0$ and an exogenous default time $\\tau$ such that under two measures the conditional survival probabilities $P(\\tau > x\\mid F_t)$ are different and no single kernel $\\gamma$ represents both simultaneously. Then the intensity is measure-dependent, the jump-size process $U$ cannot be aggregated, and Theorem 1 has no object to attach to, showing that the binding restriction is the hypothesis rather than the equation itself.","supporting_citations":[{"cited_title":"Mean-variance hedging on uncertain time horizon in a market with a jump.Applied Mathematics & Optimization, 68:413–444, 2013","cited_arxiv_id":null,"evidence_quote":"Supplies the before/after-default decomposition of erratic-horizon BSDEs that Theorem 1 extends to the second-order setting."},{"cited_title":"A note on BSDEs with singular driver coefficients","cited_arxiv_id":null,"evidence_quote":"Gives the reduction of a random-horizon BSDE to a Brownian BSDE with a singular driver, used in the auxiliary construction."},{"cited_title":"Mind the jumps: when 2BSDEs meet semi-martingales","cited_arxiv_id":"2507.01767","evidence_quote":"Provides the existence and uniqueness result for the auxiliary 2BSDE on which Lemma 3.1 relies."},{"cited_title":"Stochastic control for a class of nonlinear kernels and applications.The Annals of Probability, 46(1):551–603, 2018","cited_arxiv_id":null,"evidence_quote":"Provides the non-dominated 2BSDE framework and dynamic programming principle used throughout the paper."},{"cited_title":"What happens after a default: the conditional density approach.Stochastic processes and their applications, 120(7):1011–1032, 2010","cited_arxiv_id":null,"evidence_quote":"Supplies the conditional-density approach and the survival formula $P(\\tau>t\\mid F_t)=e^{-\\Lambda_t}$ that motivates the $P_0$-density hypothesis."},{"cited_title":"Wellposedness of second order backward SDEs","cited_arxiv_id":null,"evidence_quote":"Supplies the classical 2BSDE-to-PDE link used in Lemma 3.12 for the Markovian representation."},{"cited_title":"Dynamic programming approach to principal–agent problems.Finance and Stochastics, 22:1–37, 2018","cited_arxiv_id":null,"evidence_quote":"Supplies the weak-formulation volatility-control framework that the erratic stochastic control application extends."},{"cited_title":"Contract theory in a vuca world.SIAM Journal on Control and Optimization, 57(4):3072–3100, 2019","cited_arxiv_id":null,"evidence_quote":"Supplies the volatility-ambiguity and worst-case framework that the sup-inf application extends."}],"review_version":1}