{"id":"a125a518-1b90-4a5e-98f9-37cd6a825080","arxiv_id":"2508.20017","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A theorem showing the dual optimizer in Stretched Brownian Motion is finite almost surely under the target law and optimizing sequences converge in measure on the boundary.","lead":"The paper proves that dual optimizers in the Stretched Brownian Motion problem converge on the boundary of the target support, not just in the interior. This closes a technical gap relevant to calibration of the Bass local volatility model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof of the boundary inequality (10) explicitly omits the case ψlim(0)=∞; the gap is likely patchable, but the paper as written does not establish (10) at infinite boundary values.","rationale":"The reader's weakest_assumption focuses on the imported BBST25 off-boundary convergence theorem. That is a real dependency, but it is explicitly incorporated into the hypotheses of Theorem 1 and is external rather than a flaw in the present argument. The most load-bearing internal point is the omitted ψlim(0)=∞ case in the proof of the central inequality (10). This is an explicit proof gap and is not merely cosmetic: it concerns exactly the boundary values where the paper's new contribution is supposed to apply. I assessed whether the gap could be filled. A standard convex-analysis argument using lower semicontinuity of ψlim at 0 and uniform convergence on compact subsets of I appears to close it: any subsequence with ψn(0) bounded above would force ψn(1) to be arbitrarily large, contradicting the finite limit ψlim(1). Thus the concern does not currently invalidate the theorem, but the manuscript is not fully self-contained at this point. The reader's verdict ACCEPT is therefore unchanged, with the recommendation that the omitted case be written out explicitly. Agreement is partial because the reader mentioned the omitted case in the rationale as benign but did not identify it as the weakest assumption.","tokens_in":10506,"tokens_out":37620,"duration_ms":419314,"concrete_test":"Write out the omitted ψlim(0)=∞ case: suppose a subsequence has ψn(0)≤M. For each L, lower semicontinuity at 0 gives δ_L>0 with ψlim(x)≥L on (0,δ_L). By uniform convergence on [δ_L/2,δ_L]⊂I, for large n, ψn(x)≥L−1 on that interval. Convexity with ψn(0)≤M then forces ψn(1)≥L−1 for large n. Letting L→∞ contradicts ψn(1)→ψlim(1)<∞. If this argument is valid, the gap is benign and the theorem stands; if it fails, Theorem 1 needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of the central boundary inequality (10) in Theorem 1 is the load-bearing step: without it, the later L1 and L0 convergence arguments lack their key pointwise input. In the proof, after reducing to d=1, y=0, x0=1, the text says: 'We focus on the case ψlim(0)<∞, and leave the case ψlim(0)=∞ to the reader.' This is not a purely cosmetic omission. The conclusion ψlim∈L0(ν) permits ψlim=∞ only on a ν-null set, yet the inequality (10) must hold pointwise for every y, including those where ψlim(y)=∞. The printed convexity argument requires ψlim(0)<∞ to obtain the finite lower bound in (14); when ψlim(0)=∞, the lower-semicontinuity hypothesis gives no finite anchor, and the slope argument as written cannot be applied directly. If this omitted case were genuinely false, the central claim would fail. The case can probably be patched by combining lower semicontinuity of ψlim at 0 with the uniform convergence of the convex functions ψn on compact subsets of I: if a subsequence had ψn(0) bounded above, then for arbitrarily large L one could force ψn(1)→∞, contradicting ψn(1)→ψlim(1)∈R. But this argument is not supplied in the manuscript, so the proof of (10) is incomplete for exactly the boundary values that the paper is designed to handle.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper strengthens the convergence theory for dual optimising sequences in the martingale Benamou–Brenier problem. For an irreducible pair μ ≤_c ν in P_2(R^d), the authors prove that, after adding affine functions, any dual optimising sequence (ψ_n) that converges pointwise to a dual optimiser ψ_lim off the relative boundary C\\I actually converges to ψ_lim in ν-measure, that ψ_lim is finite ν-a.s. (i.e. ψ_lim ∈ L^0(ν)), and that liminf_n ψ_n(y) ≥ ψ_lim(y) for every y. If additionally spt(μ) is compactly contained in I, the convergence is upgraded to L^1(ν). The proof proceeds by first establishing the pointwise boundary inequality (10) via a one-dimensional convexity argument, then proving L^1 convergence under a compact-support condition using an approximation lemma (Lemma 2), and finally removing that condition by a localisation argument based on Lemmas 3 and 4.","tokens_in":10842,"tokens_out":7299,"duration_ms":85638,"significance":"If the result is correct, it resolves a genuine boundary pathology in the stretched Brownian motion / martingale Benamou–Brenier theory: the limiting dual optimizer, which was previously known to be finite on the relative interior, is now shown to be finite ν-a.s. and to be attained as the ν-measure limit of any dual optimising sequence. The proof is a coherent sequence of reductions: the auxiliary Lemmas 2–4 are natural and are stated with explicit hypotheses, and the heavy lifting is cleanly delegated to the prior parameter-free results of BBST25. The paper contains no fitted constants and the main claim is a concrete, falsifiable statement about convergence in measure. The main issue is an explicitly omitted case in the proof of the central inequality, which makes the current manuscript incomplete even though the gap appears patchable.","major_comments":[{"comment":"The proof of the boundary inequality (10) is incomplete as written. After reducing to d=1, y=0, x0=1, the text states: 'We focus on the case ψ_lim(0) < ∞, and leave the case ψ_lim(0) = ∞ to the reader.' This is not a cosmetic omission: the convexity argument requires the finite lower bound in (14), which is obtained from ψ_lim(0) < ∞. When ψ_lim(0) = ∞, lower semicontinuity gives no finite anchor at 0, so the displayed slope argument does not apply. Since inequality (10) is used pointwise for all y and is the key input for the later negative-part and L^1/L^0 convergence arguments, this gap is load-bearing. It is likely patchable, for instance by combining lower semicontinuity of ψ_lim at 0 with boundedness of ψ_n(1) and uniform convergence on compact subsets of I, but the manuscript does not supply such an argument. The omitted case must be supplied before the proof of Theorem 1 is compl","section":null},{"comment":"In the proof that ψ_lim ∈ L^0(ν) and ψ_n → ψ_lim in L^0(ν), the authors define μ_j := μ(·|K_j) for the sets K_j from Lemma 2, where K_j ⋐ I. This requires μ(K_j) > 0. However, μ may a priori charge the relative boundary C\\I, in which case some or all of the μ_j are undefined. The manuscript does not justify μ(K_j) > 0. The missing fact is presumably that irreducibility, via [BBST25, Cor. 7.7] (π_x^SBM ∼ ν for μ-a.e. x), forces μ(I) = 1: if x ∉ I, a probability π_x supported on C with mean x must be supported on a proper face, so it cannot be equivalent to ν. Since K_j ↑ I and μ(I)=1, μ(K_j)>0 for all sufficiently large j. This is a local missing justification and likely fixable, but it is needed for the localisation argument.","section":null}],"minor_comments":[{"comment":"The symbol M_2(p) is used in the proof of Lemma 2 without being defined; it should be introduced as the second moment of p.","section":null},{"comment":"The author name appears as 'W alter Schachermayer' with an unwanted space; this is a typesetting typo.","section":null},{"comment":"The statement of Lemma 4 item 2 defines ν_j ∈ P_1(R^d) but the surrounding text works with P_2 moments; this is harmless but should be made consistent.","section":null},{"comment":"The line 'By [BBST25, Lemma 7.9] and Fatou’s lemma 0 ≤ A := ... < ∞' would benefit from a one-sentence explanation of why the relevant integrals are finite, since the finiteness of A is used to justify the integrability of ψ_lim with respect to ν.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid and useful contribution, and the central claim is plausible. The explicit omission of the ψ_lim(0)=∞ case in the proof of (10) is a genuine gap in a load-bearing step, and the μ(K_j)>0 issue in the L^0 proof also needs a short justification. Both appear patchable within the scope of the paper, so I would not recommend rejection, but the manuscript should not be accepted in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Schachermayer and Siorpaes close the boundary gap left open in BBST25: for irreducible µ ≤_c ν, the dual optimiser ψ_lim is ν-a.s. finite and any dual optimising sequence converges to it in ν-measure, with an L1 upgrade when spt(µ) is compactly contained in the relative interior. The result is real and the proof is essentially right.\n\nThe new content is Theorem 1 plus Lemmas 2–4. Lemma 2's nested compact approximation via stopped SBM is a nice tool; Lemma 3's localisation by conditioning on a subset is clean; Lemma 4's reduction of L0 convergence to local pieces is straightforward but useful. The L1 proof uses dominated convergence and Fatou in the right places. The heavy dependence on BBST25 is legitimate: the off-boundary pointwise convergence and the existence/uniqueness of ψ_lim are genuine imports, not the target result.\n\nThe main soft spot is exactly what the stress-test flags. The proof of (10) says 'we focus on ψ_lim(0)<∞, and leave the case ψ_lim(0)=∞ to the reader.' That is a real omission in the written proof. But it is a minor gap, not a load-bearing one. If ψ_lim(0)=∞, lower semicontinuity forces ψ_lim(x)→∞ as x↓0, and pointwise convergence on (0,δ) then forces ψ_n(x)→∞ for each fixed small x. Convexity with ψ_n(1) converging to a finite limit then forces ψ_n(0)→∞; otherwise ψ_n(x) would stay bounded above. So (10) holds trivially in that case. The authors should write that out, but the argument is immediate from what they already have. The stress-test's suggested patch works, and the omission does not threaten the theorem.\n\nOne other thing: the assumption that ψ_n≥0 and converges pointwise on I∪C^c is imported from BBST25 and stated as 'w.l.o.g.'; that is fine given the predecessor result. No questionable data or invented entities; no fitting. Citation pattern is self-heavy but appropriate.\n\nWho should read this: anyone working on Bass martingales, stretched Brownian motion, or the convergence theory of MOT duals. It deserves a serious referee and should be accepted after a minor revision that fills the omitted case.\n\nRecommendation: send to peer review; accept with minor revision.","headline":"Closes the boundary gap in BBST25 with a sound, if slightly incomplete, proof; the omitted infinite-boundary case is a minor gap, not a real flaw.","tokens_in":11353,"tokens_out":2921,"would_cite":true,"duration_ms":32062,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G42","60G44","91G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Theorem 1 proves that for irreducible pairs in convex order, any dual optimising sequence converges to the dual optimizer in ν-measure on the whole space, including the relative boundary of the target's convex hull, and the optimizer is ν-a","keywords":["martingale optimal transport","stretched Brownian motion","Bass martingale","dual optimising sequence","convex order","irreducible pair","convergence in measure","Benamou–Brenier"],"falsifier":"Compute the dual for an irreducible pair in R^2 where ν charges the relative boundary (for example, µ uniform on a segment in the interior and ν supported on two opposite boundary points plus interior mass). If any dual optimising sequence — after affine shifts satisfying the assumed interior convergence — fails to converge in ν-measure on the boundary, or if ψ_lim is infinite on a ν-positive boundary set, Theorem 1 is refuted. The theorem predicts neither failure can occur.","tokens_in":10394,"feed_emoji":"📈","tokens_out":6661,"duration_ms":72292,"temperature":0.7,"pith_summary":"This paper closes the last gap in the convergence theory for the dual of the martingale Benamou–Brenier problem. For an irreducible pair of probability measures in convex order, prior work showed that any dual optimising sequence can be shifted by affine functions so that it converges pointwise to the dual optimizer everywhere except possibly on the relative boundary of the convex hull of the target's support. The authors prove that on that boundary the boundary pathologies do not occur: the optimizer is finite almost surely under the target measure, the shifted optimizing sequence converges to it in measure, and the pointwise liminf inequality holds at every point. If the source measure is compactly contained in the interior, the convergence improves to L1 under the target measure. The proof combines a one-dimensional convexity slope argument with a localisation procedure that conditions the stretched Brownian motion on compact interior sets.","feed_headline":"Dual optimising sequences converge in measure on the boundary","feed_subtitle":"The dual optimizer is finite almost surely and every optimizing sequence converges to it on the relative boundary.","key_machinery":"The carrying object is the dual functional D(ψ) = ∫(∫ψ dπ_x − φ_ψ(x)) µ(dx), where φ_ψ(x) is the infimum over martingale kernels of ∫ψ dp − MCov(p,γ) and MCov is maximal covariance with the standard Gaussian. The new mechanism is a one-dimensional convexity argument: any boundary point y is joined to an interior point x_0, convex functions are restricted to the segment, and the slope comparison forces the liminf inequality (10). This is supported by three localisation lemmas: Lemma 2 approximates SBM kernels by kernels supported on compact convex subsets of the interior; Lemma 3 shows conditioning µ on such a compact set preserves the stretched Brownian motion, the dual optimising property,","core_discovery":"For an irreducible pair µ ≤_c ν in P2(R^d), any dual optimising sequence (ψ_n) and dual optimizer ψ_lim satisfying the standing assumptions admit the following strengthening: ψ_lim is finite ν-a.s., (ψ_n) converges to ψ_lim in ν-measure, and the pointwise inequality liminf_n ψ_n(y) ≥ ψ_lim(y) holds for every y. If spt(µ) is compactly contained in the relative interior I, then ψ_lim ∈ L1(ν) and the convergence is in L1(ν). This removes the boundary exception from the earlier pointwise convergence result, so the full convergence picture becomes: pointwise on the interior and outside the hull, and in measure on the relative boundary where ν may charge positive mass.","pith_inferences":["Because convergence in ν-measure implies the existence of a subsequence converging ν-a.s., Theorem 1 yields an almost-everywhere convergent subsequence of any dual optimising sequence on the boundary; this direct corollary is not stated in the paper.","The L1 statement may extend beyond the compact-support condition spt(µ) ⋐ I: the proof uses compactness only to bound ψ_lim on spt(µ), so a finite ∫ψ_lim dµ could plausibly replace it.","The boundary finiteness result should stabilise numerical Bass-model calibration when the target measure charges boundary points, since the dual potential that algorithms iterate on no longer blows up on ν-positive boundary sets."],"forward_implications":["The dual convergence picture is now complete: pointwise on the interior and outside the convex hull, and convergence in ν-measure on the relative boundary.","The dual optimizer ψ_lim is ν-a.s. finite, so it is a genuine finite-valued function on the support of the target measure, not merely an extended-valued potential.","The pointwise liminf bound holds at all points, so no dual optimising sequence can dip below the limit even on boundary points.","When spt(µ) is compactly contained in the relative interior, the convergence is in L1(ν), giving the integrability and moment control needed for pricing and calibration applications.","The localisation Lemma 3 transfers dual optimality to conditioned pairs, making the measure-convergence result stable under conditioning on compact interior sets."],"supporting_citations":[{"why":"Supplies the prior result that any dual optimising sequence can be shifted by affine functions to converge pointwise to the dual optimizer on I ∪ C^c; this is the starting assumption of Theorem 1.","marker":"[BBST25, Theorems 7.8 and 7.20]"},{"why":"Gives existence of the lower semicontinuous dual optimizer ψ_lim under irreducibility and the condition µ(ri(ψ_lim < ∞)) = 1.","marker":"[BBST25, Theorem 7.6]"},{"why":"Shows the SBM kernel πSBM_x is equivalent to ν for µ-a.e. x, used in Lemma 3 to prove ν_B ∼ ν.","marker":"[BBST25, Corollary 7.7]"},{"why":"Characterises the Stretched Brownian Motion as the unique maximiser of the covariance functional, used in Lemma 3 to identify the conditioned SBM.","marker":"[BBST25, Theorem 3.3]"},{"why":"Provides the finite integral A involving ψ_lim and the SBM kernel that starts the L1(ν) argument.","marker":"[BBST25, Lemma 7.9]"},{"why":"Gives uniform convergence of convex functions on compact sets, needed to pass integrals in the proof of inequality (15).","marker":"[HUL01, Theorem 3.1.4]"},{"why":"Converts weak convergence plus second-moment convergence into W2 convergence in Lemma 2.","marker":"[Vil03, Theorem 7.12]"}],"fun_headline_variants":["Stretched Brownian motion: dual optimisers converge on boundary","Boundary exception removed: dual sequences converge in measure","Dual optimiser finite a.s., sequences converge in measure on boundary","Full convergence for stretched Brownian motion dual sequences","Compact support yields L1 convergence for dual optimisers"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is the previously established off-boundary convergence theorem: after adding affine functions, any dual optimising sequence converges pointwise to ψ_lim on the interior and outside the convex hull; without that theorem the one-dimensional slope argument cannot be applied to boundary points.","fun_headline_variants_meta":{"raw":{"variants":["Stretched Brownian motion: dual optimisers converge on boundary","Boundary exception removed: dual sequences converge in measure","Dual optimiser finite a.s., sequences converge in measure on boundary","Full convergence for stretched Brownian motion dual sequences","Compact support yields L1 convergence for dual optimisers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000367,"raw_usage":{"total_tokens":1798,"prompt_tokens":726,"completion_tokens":1072,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":991}},"tokens_in":470,"tokens_out":1072,"duration_ms":9955,"temperature":1.0,"reasoning_tokens":991,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:13:21.063969+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the dual for an irreducible pair in R^2 where ν charges the relative boundary (for example, µ uniform on a segment in the interior and ν supported on two opposite boundary points plus interior mass). If any dual optimising sequence — after affine shifts satisfying the assumed interior convergence — fails to converge in ν-measure on the boundary, or if ψ_lim is infinite on a ν-positive boundary set, Theorem 1 is refuted. The theorem predicts neither failure can occur.","supporting_citations":[],"review_version":1}