{"id":"14a05d64-249d-40e0-b289-c54824eb177d","arxiv_id":"2608.05198","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Gaussian volatility-surface models cannot stay inside the static no-arbitrage set: at an active constraint they exit with probability 1/2, so exact models must be reflected, tangent, or confined to parameterized arbitrage-free manifolds.","lead":"This paper builds a unified mathematical theory of implied and local volatility surfaces as constrained infinite-dimensional fields, and proves that Gaussian field dynamics exit the no-arbitrage set with probability tending to one half at an active constraint. It then lays out three viable construction templates, spectral hedging, neural-operator and normalizing-flow generators, and a pre-registered empirical protocol.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Half-law is proved only for window-averaged active constraints; the central claim's 'any active constraint' is broader and leaves pointwise/isolated butterfly contacts outside the H^2 statement.","rationale":"The half-law proof itself is standard and internally consistent: first-order expansion with C^1 window functionals, division by sqrt(Delta), Slutsky, and Gaussian anti-concentration. The concern is scope, not soundness. The reader's compact-window worry is real but less central to the theorem: adding wing constraints only enlarges the violation event, so it would strengthen the lower-bound conclusion rather than threaten the half-law. The more load-bearing mismatch is between the claim's 'any active constraint' and the window-active hypothesis of Theorem 7.3, which the paper itself flags in Remark 7.2. Because the limitation is explicit and the verdict is already CONDITIONAL, I would not change the verdict; the conditional label already captures the need to verify whether the 1/2 law transfers to pointwise and localized-boundary activity.","tokens_in":32759,"tokens_out":16025,"duration_ms":165957,"concrete_test":"Take a finite-rank Gaussian KL model with q smooth basis functions on a 41x21 grid; set w0 so g[w0]=0 at exactly one node and strictly positive elsewhere, with the shock covariance giving positive variance to the pointwise butterfly normal Dg_node[w0]. Estimate P(g_node(w0 + sqrt(Delta)xi + Delta mu) < 0) for Delta = 10^-2 down to 10^-6. If the limit is not 1/2, Theorem 7.3 does not extend to isolated contacts and the central claim is overbroad. Repeat with the active set an interval of width epsilon shrinking to zero to probe the positive-measure boundary of the window-functional assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 7.3 proves P(Phi(w_Delta)<0) -> 1/2 for the smoothed functionals Phi of (30), so it applies only when the constraint is active on a set of positive measure inside the window. The paper itself concedes in Remark 7.2 that an isolated point contact is invisible to the L^2 order structure of H^2; for such contacts the theorem gives no 1/2 law. The abstract and the reader's strongest_claim, however, are phrased as 'at any surface with an active constraint' and 'no unconstrained Gaussian or PCA-style surface model can be exactly statically arbitrage-free at the boundary'. In practice, butterfly constraints bind on small sets (a strike, a short-dated wing), and the finite-grid Theorem 7.6 supplies a tangency and inward-drift condition rather than an exit probability, so it does not fill the gap. A secondary scope restriction is that Corollary 7.5 is proved for the additive-noise SDE class (28); state-dependent Gaussian-type factor models are outside the proof. Neither issue makes the proved half-law false, but the advertised scope is wider than the theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a unified infinite-dimensional state-space framework for implied total variance surfaces, viewed as H^2_rho(D)-valued random fields constrained by positivity, calendar monotonicity, and a butterfly differential inequality. It establishes topological and geometric properties of the admissible set (closedness, convexity in price coordinates, nonconvexity in total-variance coordinates, tangent cones), proves a boundary incompatibility theorem: at an active constraint detected by window-averaged functionals, a nondegenerate Gaussian shock exits the admissible set with probability tending to 1/2 as the time step shrinks (Theorem 7.3), with a continuous-time trichotomy (degenerate normal noise, reflection, or submanifold confinement) in Corollary 7.5. It then develops Hilbert-space dynamics with exact modal reduction, Karhunen-Loève expansions, a vega-field hedging theory with a closed-form minimum-variance hedge, a local-volatility chart via the Dupire ratio, three viable model constructions (reflected price dynamics, tangent-projected innovations, parameter-map/submanifold dynamics), neural-operator heads, arbitrage-free normalizing flows, fading-signature fields, and a pre-registered empirical protocol with nine models, six metric families, and seven falsifiable hypotheses.","tokens_in":33063,"tokens_out":5117,"duration_ms":50249,"significance":"If the central claims hold, the paper supplies a rigorous geometric foundation for volatility-surface modeling in which static no-arbitrage is a constraint-set property rather than an afterthought. The half-law in Theorem 7.3 is a clean and honest result: the proof via Taylor expansion plus Slutsky is elementary and appears correct, and the continuous-time trichotomy is a standard and well-executed Dambis-Dubins-Schwarz/LIL argument. The paper is unusually transparent: every formal result is tiered [Proved]/[Conditional]/[Conjectural], the nonconvexity certificate in Proposition 3.3 is verified with interval arithmetic and ships with code, and Section 15 is a genuinely pre-registered design with explicit falsification conditions. The hedging formula (53) and the modal truncation error (46) are concrete, usable outputs. The main weakness is that the headline scope—'at any surface with an active constraint'—is wider than the proved statement, which covers only positive-measure active windows in H^2; this gap affects the abstract, the introduction, and the empirical hypothesis H1, and requires either a proof extension or a scope correction.","major_comments":[{"comment":"Theorem 7.3 establishes the 1/2 exit limit only for the window-averaged functionals (30), i.e., when the constraint is active on a set of positive measure inside the window. The abstract's 'at any surface with an active constraint' and the introduction's 'Gaussian field models ... are incompatible with the boundary' are broader, and Remark 7.2 itself concedes that an isolated point contact is invisible to the L^2 order structure. Since butterfly constraints in practice often bind on small sets (a single strike or short-dated wing), the gap is load-bearing. The finite-grid Theorem 7.6 gives only tangency and inward-drift conditions, not an exit probability, so it does not fill the gap. The paper should either restrict the advertised claim to 'positive-measure active window' or promote the asserted pointwise version (for H^s with s>3) from a remark to a proved theorem.","section":"§7.3, Theorem 7.3; Remark 7.2; Abstract; §1"},{"comment":"Corollary 7.5 is proved for the additive-noise equation (28), d w_t = μ_t dt + Σ dW_t. The trichotomy is then stated in the abstract and Section 1 without this restriction, but state-dependent Gaussian-type factor models d w_t = μ_t dt + B(w_t) dW_t, which are the natural lift of PCA-style models after learning, are outside the proof. The same argument plausibly extends to Lipschitz B(w), but the extension is not written down; as it stands, the advertised 'exact invariance therefore requires degeneracy, reflection, or confinement' overstates the proved scope. The proof should be extended or the statement qualified to the additive class.","section":"§7.1 and Corollary 7.5"},{"comment":"Hypothesis H1 predicts that the unconstrained Gaussian model's one-step violation frequency 'follows the standardized-margin curve ... increasing toward 0.5 as the conditioning margin and simulation step shrink.' This prediction is based on Theorem 7.3, but that theorem does not apply to pointwise or isolated active constraints on the empirical grid, where Theorem 7.6 only yields tangency and drift conditions. As written, H1 may test a stronger statement than the proved half-law supports. The design should either condition on positive-measure active regions (e.g., using window functionals on a neighborhood of the margin state) or explicitly formulate a separate, weaker prediction for isolated contacts.","section":"§15.4, Hypothesis H1"},{"comment":"The paper repeatedly justifies the compact window D=[k_min,k_max]×[τ_min,τ_max] by stating that Lee wing constraints are 'dominated by the butterfly constraint on this window' and are omitted from K. This dominance claim is asserted but not demonstrated; a counterexample on the boundary of D or a brief argument (or reference) showing that any surface satisfying the butterfly inequality on D can be extended to an arbitrage-free surface on the full half-plane would make the scope of the 'arbitrage-free' labels precise. As it stands, the skeptic's concern that finite-window feasibility may not transfer outside D is left open.","section":"§3.1 and §16"}],"minor_comments":[{"comment":"The equality T_K(w)=L(w) is tiered [Conditional] on a Robinson-type constraint qualification with the pullback of a Slater direction through B^{-1}. Since this equality is not used in the proof of Theorem 7.3, the conditional tier is acceptable, but a one-line example of a w where the pullback Slater direction is known to work would make the remark more useful.","section":"§4.2, Remark 4.3"},{"comment":"The sentence 'On the compact window D they are dominated by the butterfly constraint' would benefit from a pointer to the precise sense of domination (e.g., pointwise bound on the Lee moment slope in terms of g), since it is the only place where the omission of Lee wings is justified.","section":"§3.1"},{"comment":"The paper is careful that w_kk only exists in L^2, but the butterfly functional g is then only an L^2 element; the notation g[w]≥0 a.e. is clear, though the text could state explicitly that pointwise positivity of g is not claimed for general w∈H.","section":"§2.2, Proposition 2.1(iii)"},{"comment":"The proof of (19) divides two identities and assumes the denominator is nonzero; the subsequent Corollary 5.2 treats the zero-pole geometry, but the theorem statement could explicitly include the assumption ∂_τ w>0 and g[w]>0 to avoid ambiguity.","section":"§5, Theorem 5.1"},{"comment":"The discussion of quasi-invariance and the Ramer framework is accurate and appropriately hedged; however, the phrase 'generic coordinatewise neural flows need not satisfy these conditions' could be strengthened by a concrete example (e.g., a nonlinear shift map on ℓ^2 that sends the reference Gaussian to a singular measure) to make the warning more tangible.","section":"§13.3"},{"comment":"The paper is very long and contains several notational transitions (w, c, a, S, etc.). A consolidated notation table near the end of Section 2 would help the reader navigate the coordinate changes, especially the duality between the variance chart and the price chart.","section":"Throughout"},{"comment":"The citations to 'Noguer i Alonso 2026a,b' are to unpublished working papers by the author; in a journal version, these should be marked as forthcoming or available online, and the text should clarify their role (they are used for covariance repair taxonomy and market-mode analysis, not for the main proofs).","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a sprawling synthesis; the core contribution is Section 7, and it is essentially correct within its stated H^2 window-averaged scope. The major issue is scope alignment: the abstract and introduction claim more than Theorem 7.3 proves, and Hypothesis H1 inherits that overreach. I would not reject: the half-law and the trichotomy are novel and defensible, and the paper's tiering is a model of scholarly transparency. However, the author should either extend the theorem to pointwise contacts (under H^s, s>3) or rewrite the headline claims; both are feasible within the manuscript's scope. I also suggest the editor ask the author to trim or clearly mark the many forward-looking 'program' sections, since the empirical and some constructional parts are pre-registered designs rather than results, and the paper's length currently obscures the path from the main theorem to its applications."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: the boundary half-law is real and new. Theorem 7.3 plus Corollary 7.5 show that a nondegenerate Gaussian field shock at an active smoothed constraint exits the admissible set with probability tending to 1/2, and continuous-time exact invariance forces degenerate normal noise, reflection, or confinement. The proof is a short Taylor/Slutsky argument plus a Dambis-Dubins-Schwarz/LIL argument, both standard and locally correct. This is a genuine structural result, not a repackaging. It does not follow from Filipovic or the Gaussian PCA program.\n\nThe paper earns credit for several supporting pieces: clear state-space setup in H^2, exact modal reduction with truncation error, the vega-field hedge theorem, and an unusually honest tiering of proved/conditional/conjectural claims. The empirical section is framed as a pre-registered protocol, and no fitted constants enter the main theorem. That counts for a lot.\n\nSoft spots, in proportion:\n\n- The stress-test concern is right. Theorem 7.3 proves the 1/2 law for window functionals, i.e., constraints active on a set of positive measure. An isolated point contact is invisible in H^2; the paper says so in Remark 7.2. The abstract and intro phrase it as “any active constraint,” which is broader than the theorem. On real surfaces butterfly constraints often bind on small sets. So the abstract needs a scope sentence, but the theorem itself is fine.\n- Construction I (reflected price dynamics) is tiered Conditional, and the tangent-projection limit in Construction II is explicitly conjectural. Those are honest labels, but they mean the “three viable constructions” are not all theorems. Fine as a roadmap, not a finished theory.\n- The promised interval verifier and reproducibility checklist are referenced but not shipped. Minor in a theory paper, but the paper advertises them.\n- The compact-window assumption is explicit and reasonable, though Lee-wing constraints are asserted to be dominated without detailed argument. That struck me as a small gap, not a fatal one.\n\nBottom line: if you work on volatility surface models, this is worth your time. The half-law deserves a serious referee and should be published after the scope of Theorem 7.3 is made honest in the abstract. I’d take it for peer review.","headline":"A real and new boundary theorem for Gaussian volatility-surface models, slightly over-advertised in the abstract; deserves serious peer review, not desk rejection.","tokens_in":33530,"tokens_out":1821,"would_cite":true,"duration_ms":18386,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91G20","60H15","46E35","91G80"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a nondegenerate Gaussian shock at an active volatility-surface constraint exits the no-arbitrage set with probability 1/2, so exact static arbitrage-freeness forces degenerate noise, reflection, or submanifold…","keywords":["implied volatility surface","static arbitrage","local volatility","constrained stochastic fields","Gaussian dynamics boundary theorem","neural operators","normalizing flows","Karhunen–Loève expansion"],"falsifier":"Compute, on real option data, the one-step violation frequency of an unconstrained Gaussian KL generator conditional on small pre-step butterfly margin as the time step shrinks: the theorem predicts convergence to $1/2$ along the curve $N(\\Phi/(\\sqrt{\\Delta}\\sigma_\\Phi))$, so a frequency that stays near zero, or a finite counterexample surface with a nondegenerate Gaussian direction that never exits $K_m$, would falsify it.","tokens_in":32551,"feed_emoji":"📉","tokens_out":12190,"duration_ms":100006,"temperature":0.7,"pith_summary":"This paper treats the implied volatility surface as a single infinite-dimensional state—total variance $w_t(k,\\tau)=\\tau\\sigma^2_t(k,\\tau)$—and asks what it means for that state to stay inside the static no-arbitrage set. It proves a boundary incompatibility theorem: at any surface where a constraint (positivity, calendar monotonicity, or butterfly positivity) is active, a nondegenerate Gaussian shock exits the admissible set with probability tending to $1/2$ as the time step shrinks, and in continuous time exact invariance forces degenerate normal noise, reflection, or confinement to an arbitrage-free submanifold. The same geometry yields practical tools: Dupire local variance is the ratio $\\partial_\\tau w/g[w]$, a Karhunen–Loève expansion with exact truncation error, a closed-form minimum-variance vega-field hedge, and three viable model templates (reflected price dynamics, tangent-projected innovations, parameter maps). The result matters because Gaussian factor/PCA surface models, the standard descriptive picture since the late 1990s, cannot be exactly arbitrage-free at the boundary; model design must build the constraints into the architecture.","feed_headline":"Gaussian surface models fail at the arbitrage wall half the time","feed_subtitle":"Boundary theorem: exact no-arbitrage forces tangency, reflection, or a constrained manifold—three viable model designs follow.","key_machinery":"The load-bearing object is the total-variance field $w_t(k,\\tau)=\\tau\\sigma^2_t(k,\\tau)$ as an element of a weighted Sobolev space $H=H^2_\\rho(D)$, evolving under an $H$-valued Itô equation $dw_t=\\mu_t\\,dt+\\Sigma\\,dW_t$. Its admissible set $K_m$ is cut out by three constraint families: positivity $w\\ge m$, calendar monotonicity $\\partial_\\tau w\\ge 0$, and the Gatheral–Jacquier butterfly inequality $g[w]\\ge 0$, where $g[w]=\\bigl(1-\\frac{k w_k}{2w}\\bigr)^2-\\frac{w_k^2}{4}\\bigl(\\frac1w+\\frac14\\bigr)+\\frac{w_{kk}}2$. Two mechanisms carry the argument. First, the boundary-layer scaling: a Gaussian shock has normal component of order $\\sqrt{\\Delta}$ and drift of order $\\Delta$, so at an active constraint the signed violation $\\Phi(w_0+\\sqrt{\\Delta}\\xi)$ collapses to a centered normal, giving the $1/2$ exit probability; the continuous-time trichotomy follows from the Itô formula together with Dambis–Dubins–Schwarz and the Brownian law of the iterated logarithm. Second, the chart geometry: the pointwise Black–Scholes map $B$ makes the price set $K_{c,m,M}$ closed and convex while $K_m$ is nonconvex in total-variance coordinates, and the ratio $a=\\partial_\\tau w/g[w]$ turns the constraint functionals into the numerator and denominator of Dupire local variance.","core_discovery":"The paper's central discovery is an incompatibility between Gaussian dynamics and the boundary of the static no-arbitrage set $K_m=\\{w\\ge m,\\ \\partial_\\tau w\\ge 0,\\ g[w]\\ge 0\\}$. Theorem 7.3 shows that for any smoothed constraint functional $\\Phi$ that is active at $w_0$ and any trace-class Gaussian increment with nonzero variance along the constraint normal, the Euler shock $w_0+\\Delta\\mu+\\sqrt{\\Delta}\\xi$ violates $\\Phi$ with probability tending to $1/2$ as $\\Delta\\to 0$, independent of drift, because normal noise is order $\\sqrt{\\Delta}$ while drift is order $\\Delta$. Corollary 7.5 sharpens this to continuous time: an exactly invariant $K_m$-valued Gaussian field must have vanishing normal noise at every active constraint, so the only exits from the Gaussian class are degenerate diffusion, reflection, or evolution on a submanifold contained in $K_m$. Complementing the negative theorem, the interior is exponentially safe (Borell–TIS bound), the price chart is convex while the total-variance chart is not, and the Dupire identity $a=\\partial_\\tau w/g[w]$ makes the same two constraint functionals observable through local volatility. The paper claims this geometry reorganizes the model universe into three viable constructions and makes no-arbitrage an architectural requirement for learned generators.","pith_inferences":["The same boundary-layer argument should apply to any stochastic field with smooth equality constraints and nondegenerate Gaussian noise—yield curves, survival curves, or density surfaces—so the trichotomy of degeneracy, reflection, or submanifold confinement is likely a general template for shape-constrained infinite-dimensional dynamics.","The ratio $a=\\partial_\\tau w/g[w]$ suggests a direct diagnostic: on days with extreme local volatility, checking whether the numerator or denominator approaches zero attributes the behavior to calendar versus butterfly activity, and this diagnostic is only tested indirectly in the paper's empirical design.","For model comparison, violation frequency conditional on pre-step margin—not unconditional violation rate—is the metric that separates Gaussian from constrained generators, because the paper's interior survival bound shows both types coincide away from the boundary."],"forward_implications":["No unconstrained Gaussian or PCA-style surface model, however the drift is chosen, can be exactly statically arbitrage-free at an active constraint; the exit probability tends to $1/2$ as the time step shrinks.","Any exactly arbitrage-free surface model must implement one of three mechanisms: reflected dynamics in the convex price chart, tangent projection of Gaussian proposals, or diffusion on a parameterized submanifold whose image lies in $K_m$.","On a fixed quote grid, normally reflected price-coordinate dynamics have a unique strong solution and the projected Euler scheme converges, so static constraints can be enforced exactly pathwise.","Learned generators inherit the half-law if their noise is Gaussian with nonvanishing constraint-normal component; feasibility must be architectural, for example through simplex heads in price coordinates or cone heads in local-volatility coordinates.","The minimum-variance vega-field hedge has the closed form $\\alpha^*=(H^*CH)^{-1}H^*C\\nu$, with the increment covariance operator as the metric, and its residual variance is a computable audit statistic."],"supporting_citations":[{"why":"Supplies the butterfly functional $g[w]$, the density identity, and the SSVI conditions used to define the admissible set and Construction III.","marker":"Gatheral and Jacquier (2014)"},{"why":"Gives the characterization of arbitrage-free call-price surfaces that underlies the convex price chart.","marker":"Carr and Madan (2005)"},{"why":"Establishes the range of traded option prices used among the static constraint families.","marker":"Davis and Hobson (2007)"},{"why":"Provides the arbitrage-free implied-surface characterization that anchors the admissible set.","marker":"Roper (2010)"},{"why":"Supplies wing moment bounds; cited to justify omitting Lee constraints on the compact window and to discipline extrapolation.","marker":"Lee (2004)"},{"why":"Supplies the Hilbert-space SDE framework, trace-class covariance, and the Itô formula used in the boundary theorem and modal reduction.","marker":"Da Prato and Zabczyk (2014)"},{"why":"Provides Dambis–Dubins–Schwarz time change and the Brownian law of the iterated logarithm used in the continuous-time trichotomy proof.","marker":"Revuz and Yor (1999)"},{"why":"Gives the Gaussian concentration inequality behind the interior survival estimate.","marker":"Borell (1975)"},{"why":"Underlies the local-volatility construction from which the identity $a=\\partial_\\tau w/g[w]$ is derived.","marker":"Dupire (1994)"}],"fun_headline_variants":["Gaussian shocks breach arbitrage walls half the time—three fixes follow","Exact no-arbitrage forces tangency, reflection, or manifold confinement","Volatility geometry shows Gaussian dynamics unfit for no-arbitrage","Arbitrage-free volatility requires one of three structures","Half the Gaussian exits violate the arbitrage boundary theorem"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument is set on a fixed compact quoted window in moneyness and maturity, with time to maturity bounded away from zero, and the paper assumes that on this window the Lee wing constraints are dominated by the butterfly constraint; if infinite-domain wing limits impose material constraints that the finite window misses, the arbitrage-free labels attached to generated surfaces on the window would not transfer to real tradable surfaces outside it.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian shocks breach arbitrage walls half the time—three fixes follow","Exact no-arbitrage forces tangency, reflection, or manifold confinement","Volatility geometry shows Gaussian dynamics unfit for no-arbitrage","Arbitrage-free volatility requires one of three structures","Half the Gaussian exits violate the arbitrage boundary theorem"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000973,"raw_usage":{"total_tokens":4244,"prompt_tokens":1165,"completion_tokens":3079,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":781,"completion_tokens_details":{"reasoning_tokens":2990}},"tokens_in":781,"tokens_out":3079,"duration_ms":21598,"temperature":1.0,"reasoning_tokens":2990,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:42:20.902227+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, on real option data, the one-step violation frequency of an unconstrained Gaussian KL generator conditional on small pre-step butterfly margin as the time step shrinks: the theorem predicts convergence to $1/2$ along the curve $N(\\Phi/(\\sqrt{\\Delta}\\sigma_\\Phi))$, so a frequency that stays near zero, or a finite counterexample surface with a nondegenerate Gaussian direction that never exits $K_m$, would falsify it.","supporting_citations":[{"cited_title":"2014 , address =","cited_arxiv_id":null,"evidence_quote":"Supplies the Hilbert-space SDE framework, trace-class covariance, and the Itô formula used in the boundary theorem and modal reduction."}],"review_version":2}