{"id":"363884ae-07e3-4053-a8be-38aac32d295c","arxiv_id":"1908.09208","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Path-dependent multidimensional forward-backward stochastic differential equations have a unique stable solution whenever a constructed decoupling field and a dominating ODE stay bounded.","lead":"This paper proves existence, uniqueness, and stability for a class of equations in which the current movement of a random process depends on its entire past path, not just its current value. The result gives a rigorous foundation for path-dependent control problems in finance and economics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 4.2's patching proof silently needs the unstated condition Kmax|∇zσ|∞<1 at every step; without it Theorem 4.1 cannot be reapplied, so the global wellposedness claim is not established as written.","rationale":"The reader's verdict of CONDITIONAL is appropriate, but the sharpest reason is not the D_t=0 degeneracy. For two distinct initial paths, D_t is positive throughout, and the paper's quotient convention handles the limiting case of coincident paths. The more serious gap is in the stepwise patching proof of Proposition 4.2: it reapplies Theorem 4.1 using only the bound L_i≤Kmax, while Theorem 4.1 requires L_i|∇zσ|∞<1. The paper neither proves this smallness condition nor supplies a substitute local argument. Proposition 4.6 being imported from the Markovian theorem [5] is a second gap, but it is downstream: even with a maximal interval in hand, the patching in Proposition 4.2 still needs the missing smallness condition. Since no demonstrated counterexample is given, the appropriate recommendation remains CONDITIONAL rather than ACCEPT or REJECT; the concerned sharpens the justification for that verdict.","tokens_in":19040,"tokens_out":27393,"duration_ms":273071,"concrete_test":"Re-derive the induction of Proposition 4.2 with explicit bookkeeping of the terminal Lipschitz constant L_i at each step: compute the maximal admissible interval length δ(L_i) from the contraction estimate in Theorem 4.1 and check whether the bound L_i ≤ Kmax is sufficient to guarantee δ(L_i)>0. A decisive linear test case is dX_t=(Y_t+Z_t)dW_t, dY_t=-f(t,X_t)dt+Z_t dW_t, Y_T=kX_T with k<1, choosing f so that the decoupling field is a(t)x and a(t) crosses 1/|∇zσ|∞: if a bounded dominating ODE exists on [0,T], Proposition 4.2 is false; if not, identify the missing estimate yielding Kmax|∇zσ|∞<1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is in the proof of Proposition 4.2. After inequality (14), the text says that Kmax dominates the Lipschitz constants and therefore one can iterate Theorem 4.1 with the same interval length ε̄. But Theorem 4.1's contraction proof only produces a positive interval when the terminal Lipschitz constant L satisfies L|∇zσ|∞<1; in the limit T→0 the contraction ratio γ(ε,T) tends to L^2|∇zσ|∞^2 plus an arbitrarily small K0ε term. From the characteristic-BSDE comparison the proof only obtains L_i^2 ≤ y_{t_i} ≤ Kmax^2. It never verifies the necessary inequality Kmax|∇zσ|∞<1. If the dominating ODE is bounded but its maximum exceeds 1/|∇zσ|∞, no ε̄ is supplied and the induction cannot start; no alternative local-existence argument is given. This gap is separate from the D_t=0 degeneracy: for two distinct initial paths D_t>0 on the whole interval, so H_t, α_t, β_t are well-defined and a limiting convention is available for coincident paths.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies wellposedness of multidimensional forward-backward SDEs whose coefficients may depend on the whole past path of the forward component. The authors introduce a decoupling random field on path space, prove local existence and uniqueness under the condition K1|∇zσ|∞<1 via a contraction argument, and then attempt to patch local decoupling fields using a characteristic BSDE and a dominating ODE. They state a global wellposedness criterion (Proposition 4.2), treat the decoupled and drift-controlled cases, introduce a maximal interval in the general case following Fromm-Imkeller, and prove a stability estimate for nearby path-dependent FBSDEs. The central claimed contribution is the extension of the decoupling-field method from Markovian to path-dependent coefficients.","tokens_in":19254,"tokens_out":7597,"duration_ms":72721,"significance":"If the technical gaps are closed, the paper would provide a useful extension of the decoupling-field framework to path-dependent FBSDEs, with natural applications to path-dependent stochastic control and principal-agent problems. The local contraction proof in Theorem 4.1 is written out in detail, and the stability theorem gives a concrete quantitative estimate of a type not present in the earlier Markovian literature. However, the global patching argument in Proposition 4.2 has a load-bearing gap, and several supporting results are either imported from the Markovian setting without adaptation or only sketched.","major_comments":[{"comment":"The patching induction is incomplete. At each step, Theorem 4.1 can only be reapplied if the Lipschitz constant of the newly constructed terminal condition satisfies L_i |∇zσ|∞ < 1. From (14) the proof obtains L_i^2 ≤ y_{t_i} ≤ Kmax^2, hence only L_i ≤ Kmax; the strict inequality Kmax |∇zσ|∞ < 1 is never proved. If the bounded solution of the dominating ODE has sup y_t larger than |∇zσ|∞^{-2}, no uniform ε̄ exists and the iteration cannot start. This is a load-bearing gap in the paper's central global-wellposedness claim.","section":"Section 4.2.1, proof of Proposition 4.2, inequalities (14)-(15)"},{"comment":"The proof is delegated to [5, Theorem 2], but the object here is path-dependent: the decoupling field is defined on C([0,T],R^d) and the gradient ∇x u is a derivative on path space. The Markovian proof in [5] does not automatically cover this setting; an adaptation showing that the same maximal-interval argument works with the path norm ||·||_{2,t} is required. Since Proposition 4.7 and the general-case wellposedness rely on Proposition 4.6, this cannot remain a citation-only step.","section":"Section 4.2.5, Proposition 4.6"},{"comment":"The quantities H_t=|Y_t|^2/D_t^2, α_t=Z_t/D_t and β_t=X_t/D_t are undefined whenever D_t=||ΔX||_{2,t}=0, for instance when the two forward paths coincide on an interval. No convention or limiting argument is given, and the Itô derivation of the characteristic BSDE requires a justification in this degenerate case. This affects every result that uses the characteristic BSDE, including Proposition 4.2.","section":"Section 4.2.1, Eq. (12) and preceding definitions"},{"comment":"The displayed Itô computation is not carried out, and the sign conclusion 'which is smaller than 0 by (21)' is not transparent; as printed, the identity has mismatched factors of H_t and D_t and does not clearly follow from condition (21). Since Theorem 4.4 supplies global wellposedness in the drift-controlled case b=b_t(x,y), σ=σ_t(x), this proof should be written out in detail.","section":"Section 4.2.3, proof of Theorem 4.4"}],"minor_comments":[{"comment":"The partition condition 'ti+1 − ti > δ' should be 'ti+1 − ti ≤ δ', since Definition 3.1 applies only to intervals of length at most δ.","section":"Section 6, proof of Theorem 3.1"},{"comment":"The display writes ∫_T^t ... dt and ∫_T^t ΔZ_s dW_s as if these were the terms of dΔY_t; the correct differential form is dΔY_t = −(... )dt + ΔZ_t dW_t. The current notation makes the proof difficult to follow.","section":"Section 5, proof of Theorem 5.4, Eqs. (32)-(33)"},{"comment":"The statement 'ess sup H_t is a Lipschitz constant of u_t' should refer to ess sup sqrt(H_t), because inequality (14) bounds |u(t,X)−u(t,X′)|^2 by y_t times the path norm, so the Lipschitz constant is the square root of the bound.","section":"Section 4.2.4, text before Theorem 4.5"},{"comment":"The right-hand side of (21) uses x_T; since g is defined on the whole path space C([0,T],R^d), please clarify whether x_T denotes the terminal value x(T) or the whole path x, and similarly for the notation Δx_t used in the same condition.","section":"Section 4.2.3, condition (21)"},{"comment":"The comparison 'H_t ≤ y_t' is invoked without a citation or a proof; since this comparison principle for quadratic BSDEs is a key ingredient of the patching argument, a reference or a short justification should be provided.","section":"Section 4.2.1, proof of Proposition 4.2, after Eq. (13)"}],"recommendation":"major_revision","confidential_remarks":"The main gap in Proposition 4.2 is the crux of the paper: either the missing condition Kmax|∇zσ|∞<1 must be added, or an alternative patching argument must be supplied. If only the added condition is used, the theorem becomes narrower but remains meaningful. The reliance on [5] for Proposition 4.6 should also be addressed directly, since the path-dependent setting is exactly the novelty claimed by the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick read of arXiv:1908.09208 (Hu). The headline: this is a credible attempt to extend the decoupling-field program to path-dependent FBSDEs, and the local wellposedness part is genuinely useful, but the global patching and maximal-interval results do not hold as written — there is a missing condition in Prop 4.2 and an imported theorem in Prop 4.6.\n\nWhat's new: the path-space norm ||x−x'||_{2,t} = (∫|x−x'|² + |x(t)−x'(t)|²)^{1/2} and the variational/characteristic BSDE on path space. Theorem 4.1's contraction proof is actually written out, not waved away, and it correctly identifies K1|∇zσ|∞<1 as the smallness condition. The stability result is a nice addition. Citation habits are fine; the one self-citation is contextual.\n\nSoft spots, in order:\n1. Proposition 4.2: the patching argument needs to reapply Theorem 4.1 at each step, which requires the Lipschitz constant of the current decoupling field to satisfy L|∇zσ|∞<1. From the dominating ODE and characteristic BSDE, the proof only gets L ≤ Kmax = sqrt(max_t y_t). It never checks Kmax|∇zσ|∞<1. If the dominating ODE's bounded solution has max above 1/|∇zσ|∞, no ε̄ exists and the induction cannot start. That is a genuine gap, not a typo. Adding (max y_t)|∇zσ|∞<1 to the statement would fix it, but then the theorem is much weaker.\n2. Proposition 4.6, which provides the maximal interval for the general case, is disposed of with \"we refer to [5, Theorem 2]\". That is a Markovian theorem; the path-dependent proof is exactly what this paper is supposed to supply. This is load-bearing.\n3. Theorem 4.4's proof has an unverified inequality and notation slips; Theorem 5.4's displayed equation for dΔY_t is garbled. These are addressable, but they make verification harder.\n\nThe D_t=0 degeneracy in the characteristic BSDE is minor; a limiting convention should work.\n\nWho this is for: stochastic analysts working on path-dependent control and contract theory. If the gaps are fixed, this becomes a solid contribution. Right now I would not cite it for the global theorems.\n\nRecommendation: send it to a serious referee — the topic and the local result justify that. But the referee should be told to demand the missing patching condition and a real proof of Prop 4.6, likely major revision.","headline":"A credible extension of the decoupling-field program to path-dependent FBSDEs, with a solid local wellposedness proof and useful stability result, but the global patching and maximal-interval theorems have real gaps that need fixing.","tokens_in":19803,"tokens_out":5376,"would_cite":false,"duration_ms":47577,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H07","60H30","35R60","34F05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A bounded dominating ODE makes path-dependent FBSDEs globally well-posed.","keywords":["Decoupling Random Field","Forward-Backward SDE","Characteristic BSDE","Backward Stochastic Riccati Equations","Stability Theorem","path-dependent FBSDE","dominating ODE","wellposedness"],"falsifier":"Find two forward solutions that agree on a nontrivial interval and differ later; if the ratio $|\\Delta Y_t|^2/\\|\\Delta X\\|_{2,t}^2$ has no bounded extension at the coincidence time for a system satisfying the hypotheses of Proposition 4.2, the comparison argument collapses and the global wellposedness criterion fails.","tokens_in":18805,"feed_emoji":"🎲","tokens_out":13453,"duration_ms":123325,"temperature":0.7,"pith_summary":"Path-dependent forward-backward SDEs — systems where the coefficients at time $t$ may depend on the whole history of the forward process — appear when solving stochastic control problems with path-dependent contracts. The paper sets out to prove that such equations are well posed on arbitrary time intervals, not just locally. Its route is a random decoupling field on path space: a function $u_t(\\omega,x)$ with $Y_t = u_t(\\omega,X_{\\cdot\\wedge t})$ along solutions. Provided a scalar dominating ODE for the associated characteristic BSDE stays bounded over $[0,T]$, local decoupling fields can be patched into one global field, giving a unique solution; a separate theorem quantifies how solutions move when the coefficients are perturbed. If true, this removes the Markovian restriction from a standard wellposedness toolkit and covers control problems whose contracts depend on the whole path.","feed_headline":"Global solutions exist for path-dependent forward-backward equations","feed_subtitle":"A decoupling-field construction extends local existence of solutions to any time horizon.","key_machinery":"The load-bearing object is the decoupling field, a progressively measurable random function $u:[0,T]\\times\\Omega\\times C([0,T],\\mathbb{R}^d)\\to\\mathbb{R}^n$ with $Y_t = u_t(X)$ along any solution; 'regular' means Lipschitz in the $L^2$-type path norm $\\|x-x'\\|_{2,t}$. The argument is carried by two auxiliary objects built from differences of two solutions: the normalized ratio $H_t = |\\Delta Y_t|^2/\\|\\Delta X\\|_{2,t}^2$ and the characteristic BSDE $dH_t = -F_t(H_t)\\,dt + N_t\\,d\\tilde W_t$, whose polynomial driver $F_t$ is bounded above by a deterministic dominating ODE. The comparison principle for quadratic BSDE transfers boundedness of the dominating ODE to boundedness of $H_t$, which is exactly a Lipschitz bound on the decoupling field; that bound lets the local existence theorem be reapplied step by step to cover all of $[0,T]$.","core_discovery":"The paper's central claim is Proposition 4.2: for a path-dependent multidimensional FBSDE satisfying the global Lipschitz Assumption 1, the existence of a continuously differentiable $G$ such that $\\dot y_t = -G(t,y_t)$ is a dominating ODE for the characteristic BSDE, together with a bounded solution of that ODE on $[0,T]$, implies a unique regular decoupling field on all of $[0,T]$; by Theorem 3.1 this is exactly wellposedness. The supporting local result is Theorem 4.1: under the contraction condition $K_1|\\nabla_z\\sigma|_\\infty<1$, a unique solution exists on a small time interval whose length depends only on the dimensions and Lipschitz constants. The paper also proves Theorem 5.4, a stability estimate showing that if $(b,\\sigma,f,g)$ are replaced by nearby coefficients, the resulting solutions differ by at most a constant times the coefficient differences, measured by $\\sup_t \\mathbb{E}[\\|\\Delta X\\|_{2,t}^2 + |\\Delta Y_t|^2 + \\int_t^T |\\Delta Z_s|^2\\,ds]$.","pith_inferences":["The stability estimate suggests a natural numerical test: approximate path-dependent coefficients by piecewise-constant or Markovian ones and use Theorem 5.4 to turn coefficient error into an explicit solution error; the paper does not perform such an experiment.","The proof's normalized ratios divide by the path distance between two forward solutions, so at moments when the two paths coincide the argument is undefined; a limiting convention could likely repair this, but none is given.","Because the dominating ODE is scalar, its explosion time could be computed numerically for a concrete control problem, making $T_{\\max}$ a quantitative horizon check before solving the FBSDE."],"forward_implications":["Every path-dependent FBSDE whose characteristic BSDE admits a bounded dominating ODE has a unique solution on $[0,T]$, regardless of how strongly the coefficients are coupled.","Decoupled path-dependent FBSDEs are well posed for every horizon $T$, because their characteristic BSDE admits a linear dominating ODE (Proposition 4.3).","For systems with $b=b_t(X,Y_t)$ and $\\sigma=\\sigma_t(X)$, the sign condition (21) on increments of the coefficients and terminal data guarantees a global unique solution.","Regularity of the decoupling field persists on the maximal interval, and if that interval is open at the left end, the product $|\\nabla_x u|\\cdot|\\nabla_z\\sigma|_\\infty$ must approach 1 there (Proposition 4.7).","Small changes in the coefficients produce small changes in the solution triple, with the explicit bound of Theorem 5.4 in the norm $\\sup_t \\mathbb{E}[\\|\\Delta X\\|_{2,t}^2 + |\\Delta Y_t|^2 + \\int_t^T |\\Delta Z_s|^2\\,ds]$."],"supporting_citations":[{"why":"supplies the decoupling-field patching template and the stability proof structure this paper extends to path-dependent coefficients.","marker":"[12]"},{"why":"defines the maximal interval and supplies the existence and regularity of the decoupling field used in Proposition 4.6.","marker":"[5]"},{"why":"extends the decoupling-field method to high-dimensional backward processes, the multidimensional comparison argument this paper follows.","marker":"[22]"},{"why":"provides the comparison principle for quadratic BSDE used to dominate the characteristic BSDE by the ODE.","marker":"[8]"},{"why":"is the local backward-forward existence result generalized by Theorem 4.1 to path-dependent coefficients.","marker":"[1]"},{"why":"supplies the a priori estimates for SDE and BSDE used in the stability section.","marker":"[3]"}],"fun_headline_variants":["Path-dependent FBSDEs: local existence extends to global","Decoupling fields yield global wellposedness for path-dependent FBSDEs","Stable solutions for path-dependent forward-backward SDEs","Patching local solutions solves path-dependent FBSDEs globally","Global wellposedness via decoupling fields for path-dependent SDEs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The patching argument assumes the ratio $H_t=|\\Delta Y_t|^2/\\|\\Delta X\\|_{2,t}^2$ stays well defined and bounded even when the denominator vanishes because the two forward paths coincide.","fun_headline_variants_meta":{"raw":{"variants":["Path-dependent FBSDEs: local existence extends to global","Decoupling fields yield global wellposedness for path-dependent FBSDEs","Stable solutions for path-dependent forward-backward SDEs","Patching local solutions solves path-dependent FBSDEs globally","Global wellposedness via decoupling fields for path-dependent SDEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1310,"prompt_tokens":925,"completion_tokens":385,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":294}},"tokens_in":541,"tokens_out":385,"duration_ms":4176,"temperature":1.0,"reasoning_tokens":294,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:18:50.701139+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find two forward solutions that agree on a nontrivial interval and differ later; if the ratio $|\\Delta Y_t|^2/\\|\\Delta X\\|_{2,t}^2$ has no bounded extension at the coincidence time for a system satisfying the hypotheses of Proposition 4.2, the comparison argument collapses and the global wellposedness criterion fails.","supporting_citations":[{"cited_title":"On well-pose dness of forward-backward SDEs–a unified approach","cited_arxiv_id":null,"evidence_quote":"supplies the decoupling-field patching template and the stability proof structure this paper extends to path-dependent coefficients."},{"cited_title":"Fromm and P","cited_arxiv_id":null,"evidence_quote":"defines the maximal interval and supplies the existence and regularity of the decoupling field used in Proposition 4.6."},{"cited_title":"The Wellposedness of FBSDEs (II)","cited_arxiv_id":"1708.05785","evidence_quote":"extends the decoupling-field method to high-dimensional backward processes, the multidimensional comparison argument this paper follows."},{"cited_title":"Backward stochastic differential equatio ns and partial differential equations with quadratic growth","cited_arxiv_id":null,"evidence_quote":"provides the comparison principle for quadratic BSDE used to dominate the characteristic BSDE by the ODE."},{"cited_title":"Backward-forward stochastic differential equa tions","cited_arxiv_id":null,"evidence_quote":"is the local backward-forward existence result generalized by Theorem 4.1 to path-dependent coefficients."},{"cited_title":"Contract Theory in Continuous-Time Models","cited_arxiv_id":null,"evidence_quote":"supplies the a priori estimates for SDE and BSDE used in the stability section."}],"review_version":1}