{"id":"1461d869-fde7-4886-bc52-5d507e264a39","arxiv_id":"2501.11314","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The optimal stopping rule for soft classification of a Brownian drift is characterized by two free boundaries, with explicit asymptotic behavior in the signal-to-noise ratio.","lead":"This paper introduces and solves a 'soft classification' version of the classic sequential testing problem for a Brownian motion's drift, where the decision is a confidence score rather than a forced yes or no. It shows the optimal rule stops when the posterior belief crosses two boundaries, and describes how those boundaries depend on the signal-to-noise ratio.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1 is not proved in the paper: the identification of the value function with the free-boundary solution is delegated to references that treat the hard-classification loss, leaving the central claim unverified.","rationale":"After reading the paper, the free-boundary construction in Section 2 is plausible and the geometry via Bisztriczky is a nice idea. The main weakness is not the (G2) assumption, which is a limitation but is assumed by the theorem; it is the unproved verification theorem. The paper's proof of Theorem 3.1 is a referral to other works, and the cited works address different loss functions. The hard-classification penalty has a kink and is not C^2, so the smooth-fit and superharmonic verification in [16, Theorem 21.1] does not automatically carry over to smooth concave g. Similarly, [13] has different structural hypotheses. The bound on |π(1−π)g'(π)| is cited from the companion paper without a proof; it is probably true for concave g with g(0)=g(1)=0, but it should be established in this paper. This gap is load-bearing because Theorem 3.1 is exactly the equivalence between the abstract optimal stopping problem and the concrete free-boundary solution; without it the paper solves a free-boundary problem but does not solve the sequential decision problem it introduced. The reader's verdict CONDITIONAL already flags this in the rationale, but the stated weakest assumption (G2) is not where I would put the load. Hence partial agreement.","tokens_in":14677,"tokens_out":17297,"duration_ms":160322,"concrete_test":"Complete the verification for the simplest non-linear case, the L1 loss g(π)=2π(1−π), by writing out the Peskir–Shiryaev argument explicitly: check that V in (19) is C^1 on (0,1) with V≤g and AV=−K^{-1} on (A*,B*); apply Dynkin's formula to V(Π_t) for the candidate τ_{A*,B*}; prove the stochastic integral term is a true martingale by computing V' in closed form and verifying E[∫_0^τ (V'(Π_s))^2 ds] < ∞ for all initial π. If this succeeds for L1 but reveals an extra condition needed for general (G1)-(G2), the theorem's scope must be narrowed; if it fails, the central claim is undermined.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that V*(π)=inf_τ Eπ[cτ+g(Πτ)] is the unique non-trivial solution to (7)-(13). Theorem 3.1 is the verification step that actually identifies V* with the free-boundary solution V from (19), yet its entire proof is a citation: 'The former claim is proved analogously to [13, Proposition 3.1]. The arguments for the latter proceed as in [16, Theorem 21.1].' Neither reference treats soft-classification losses satisfying (G1)-(G2). [16, Theorem 21.1] is formulated for the hard-classification penalty g(π)=a1π∧a2(1−π), which is not C^2 and has a kink; its proof exploits that specific structure. [13, Proposition 3.1] concerns problems with linear observation costs and different structural assumptions. The paper also borrows the estimate |π(1−π)g'(π)|≤M from the companion work [2, Remark 2.3] without proof. Thus the paper does not demonstrate that the candidate V is superharmonic, that the stopping time τ_{A*,B*} is optimal, or that the relevant stochastic integrals are true martingales for general g. All downstream results (Corollary 3.2, Theorem 3.3, Section 4 asymptotics) rest on this unproved identification. This is the most load-bearing concern because even a correct free-boundary analysis would not suffice to solve the sequential testing problem if the value function were not the constructed solution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the Bayesian sequential soft-classification problem for the drift of a Brownian motion. The observer pays a linear observation cost and a terminal loss g(Πτ) depending on the posterior probability, where g is induced by a soft-classification loss and satisfies concavity and a unimodality condition (G1)-(G2). The main results are: (i) a free-boundary problem (7)-(13) whose nontrivial solution is unique when Ag(π0) < -K^{-1}, with boundaries A*, B* characterized by the two transcendental equations (17)-(18); (ii) a verification theorem (Theorem 3.1) identifying the value function with this solution; (iii) a convex-envelope representation V* - 2K^{-1}Ψ = convex envelope of H; and (iv) asymptotic and monotonicity results for the boundaries as the signal-to-noise ratio K varies. The paper also contains numerical illustrations comparing the soft-classification boundaries with the classical hard-classification boundaries.","tokens_in":15004,"tokens_out":8998,"duration_ms":82980,"significance":"If the verification step were fully supplied, this would be a clean and useful extension of the classical Wiener sequential testing problem to soft-classification losses. The free-boundary construction, the use of Bisztriczky's theorem to prove existence and uniqueness of the common tangent, the explicit equations for the boundaries, and the asymptotic analysis are genuine contributions. The paper has no fitted parameters, and the structural results are derived rather than calibrated. However, the central claim that the value function equals the solution of the free-boundary problem is currently not established in the manuscript, and all subsequent results depend on that identification.","major_comments":[{"comment":"The proof of Theorem 3.1 is not a proof of the stated result: it consists of two citations and one cited estimate. [16, Theorem 21.1] is formulated for the hard-classification loss g(π)=a1π∧a2(1−π), which is not C² and has a kink; [13, Proposition 3.1] treats linear-cost problems under different structural assumptions. Neither establishes that the candidate V in (19) is superharmonic, that the stopping time τ_{A*,B*} is optimal among all F^Π-stopping times, or that the relevant stochastic integrals are true martingales for general g satisfying (G1)-(G2). Since Corollary 3.2, Theorem 3.3, and all of Section 4 depend on this identification, this is load-bearing. The authors should provide a self-contained verification under (G1)-(G2) or a rigorous reduction that verifies all hypotheses of the cited results.","section":"§3, Theorem 3.1"},{"comment":"The proof of the convex-envelope representation uses a reduction to 'regular' stopping times delegated to [13, Proposition 2.1] and applies Dynkin's formula to H, g, and Ψ without stating the integrability and boundedness conditions needed for the stochastic integral to be a true martingale. In particular, the estimate |π(1−π)g'(π)|≤M is merely quoted from [2, Remark 2.3]; for general concave g with g(0)=g(1)=0 this is not proved here. The claim that V*−2K^{-1}Ψ is convex also relies on the identification in Theorem 3.1. Please include the missing estimates and make the justification self-contained.","section":"§3, Theorem 3.3"},{"comment":"The proof of Proposition 4.2(i) says that the derivative of A*(K) (resp. B*(K)) 'must be positive (resp. negative)', but the displayed formulas and the stated conclusion (A* decreasing, B* increasing) require the opposite signs: dA*/dK≤0 and dB*/dK≥0. The numerator signs cited appear consistent with the stated monotonicity, so this is likely a typo, but it should be corrected.","section":"§4, Proposition 4.2(i)"}],"minor_comments":[{"comment":"The proof of Lemma 2.2 contains an unexplained constant C1 and C0 and a sign change in the double integral; as written, the displayed equality after integrating the bound is not derived. The conclusion is correct, but the argument should be rewritten for clarity.","section":"§2, Lemma 2.2"},{"comment":"In the statement of Theorem 2.1, the notation 'π∗ < π∗' should presumably be 'π_* < π^*', and the constraints should read A* ≤ π_* < π^* ≤ B*. The current typography makes the statement difficult to parse.","section":"§2, Theorem 2.1"},{"comment":"The notation [π, π∗] and [π∗, π] is overloaded: the underline and overline on π are easy to confuse with the asterisks for π_* and π^*. Please use a clearer notation, such as π_l, π_r or π_- and π_+.","section":"§2, Proposition 2.6"},{"comment":"The statements that the soft-classification boundaries are contained in the classical boundaries for small K and not for large K are based on numerical comparison; they should be labelled as numerical observations rather than proved results.","section":"§5"},{"comment":"The paper uses the symbol A both for the infinitesimal operator in (6) and for the lower stopping boundary. Although this is common, a different symbol for one of the two would improve readability.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a short note whose central verification theorem is delegated to references that do not cover the present class of losses, and a key estimate is taken from the authors' companion preprint without proof. I recommend major revision rather than rejection because the free-boundary construction appears sound and the missing verification is in principle within the scope of the paper. The authors should not rely on [2] for a load-bearing estimate in a published version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the new problem formulation is interesting and the free-boundary analysis is largely self-contained, but the proof of the main verification theorem is outsourced to references that don't handle this loss class, so the central claim is not fully established in this manuscript.\n\nWhat's actually new: the soft classification formulation with general concave penalties (cross-entropy, L1, L2), the characterization of the value function as the convex envelope of H (Theorem 3.3), the asymptotic bounds on the boundaries (Propositions 4.2–4.3), and the use of Bisztriczky's theorem to prove uniqueness of the common tangent (Lemma 2.5). The analysis of H in Lemma 2.3 is careful and the derivation of the free-boundary equations (17)–(18) is correct. There are no fitted parameters or post-hoc predictions; the paper is honest about the role of (G2).\n\nSoft spots: the big one is Theorem 3.1. Its proof is a citation to [13] and [16], neither of which treats smooth concave losses satisfying (G1)–(G2). The estimate |π(1−π)g'(π)| ≤ M is borrowed from the companion paper [2] without proof. As it stands, the paper does not demonstrate superharmonicity of the candidate V or optimality of the hitting time for general g. This is fixable but not trivial; a referee should ask for the full verification. A smaller issue: Proposition 2.6 asserts the tangency points are interior \"straightforward to check\" without giving the argument. That's minor. The paper's own conclusion (ii) correctly notes that without (G2) the continuation region can be disconnected, so the main characterization is conditional on that assumption.\n\nShould you read it? If you work on sequential testing or free-boundary problems, yes. The convex envelope representation and the boundary asymptotics are worth having. It deserves a serious referee, but I'd want the verification gap closed before treating Theorem 3.1 as established.","headline":"Interesting new soft-classification sequential testing problem with a largely self-contained free-boundary analysis, but the main verification theorem is delegated to references and should be filled in before the result is treated as fully established.","tokens_in":15505,"tokens_out":2553,"would_cite":false,"duration_ms":24569,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G35","60G40","62L10","62L15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper solves the soft-classification version of Bayesian sequential testing for the drift of a Brownian motion: the optimal policy is to stop at the first time the posterior exits an interval $(A^*,B^*)$ when the signal is strong…","keywords":["soft classification","sequential testing","optimal stopping","filtering","free-boundary problems","Brownian motion drift","information ratio"],"falsifier":"A concrete numerical check: for the L1 penalty $g(\\pi)=2\\pi(1-\\pi)$ with running cost $c=1$ and squared signal-to-noise ratio $K=7$, the theorem says $V^*\\equiv g$ and immediate stopping is optimal; a simulation-based approximation of (4) that finds any bounded stopping rule with expected cost below $g(\\pi_0)$ at some starting prior would refute the verification theorem. At $K=9$ the same simulation should reproduce the expected cost of first exit from $(A^*,B^*)$ obtained from (17)–(18).","tokens_in":14492,"feed_emoji":"🎯","tokens_out":17647,"duration_ms":151942,"temperature":0.7,"pith_summary":"In the classic Bayesian sequential test, an observer pays a running cost to watch a noisy Brownian signal and must eventually issue a hard yes/no verdict about the drift. This paper relaxes that verdict: the loss is a soft score $g(\\pi)$ charged according to the posterior probability at the stopping time, with zero penalty only when the observer is certain. The paper establishes that the optimal policy is still simple: keep watching while the posterior $\\Pi_t$ remains between two constants $A^*$ and $B^*$, stop at the first exit, and if the squared signal-to-noise ratio $K=(\\alpha/\\sigma)^2$ is too small, stop immediately with $V^*\\equiv g$. The two thresholds are the unique solution of a free-boundary problem, and the value function is shown to be the largest convex minorant of a transformed penalty. A sympathetic reader cares because it gives the soft-classification analogue of the classical hard-classification solution with the same semi-explicit structure, plus a phase transition in the information ratio that the hard problem does not have.","feed_headline":"Two thresholds solve Bayesian soft classification exactly","feed_subtitle":"Keep sampling while your belief sits between two thresholds; stop immediately when the signal is too weak.","key_machinery":"The central object is the free-boundary problem (7)–(13) together with the transformed penalty $H(\\pi)=g(\\pi)-2K^{-1}\\Psi(\\pi)$, where $\\Psi(\\pi)=(1-2\\pi)\\log(\\pi/(1-\\pi))$. Equation (7) fixes the curvature of the candidate value function inside the waiting region, equations (8)–(11) impose the value and smooth-fit conditions at the boundaries, and conditions (12)–(13) force the candidate to lie below the penalty inside the waiting region and equal it outside. The existence and uniqueness of the boundary pair is carried by a common-tangent construction: restricting $H$ to the two convex pieces on either side of its concave middle, equations (17)–(18) ask for a single line tangent to both pieces, and the classical theorem that two strictly separated convex bodies admit exactly two common tangents supplies the unique solution. The same structure yields the convex-envelope representation of the value function.","core_discovery":"On the paper's own terms, the central claim is that the value function $V^*(\\pi)=\\inf_\\tau \\mathbb{E}_\\pi[c\\tau+g(\\Pi_\\tau)]$ is the unique non-trivial $C^2((0,1)\\setminus\\{A^*,B^*\\})\\cap C^1(0,1)$ solution of the free-boundary problem (7)–(13) whenever the curvature condition $Ag(\\pi_0)<-K^{-1}$ holds. The continuation region is exactly $(A^*,B^*)$, with $(A^*,B^*)$ the unique solution pair of the consistency equations (17)–(18) subject to $A^*\\le \\pi_*<\\pi^*\\le B^*$, and the smallest optimal stopping time is the first exit from that interval. When $Ag(\\pi_0)\\ge -K^{-1}$, no observation is worthwhile: $V^*\\equiv g$. The solution also satisfies $V^*(\\pi)-2K^{-1}\\Psi(\\pi)=\\inf_\\tau\\mathbb{E}_\\pi[H(\\Pi_\\tau)]$, i.e. it is the largest convex minorant of $H=g-2K^{-1}\\Psi$ with $\\Psi(\\pi)=(1-2\\pi)\\log(\\pi/(1-\\pi))$.","pith_inferences":["An ordering left unproved by the paper is that the soft-classification boundaries cross the hard-classification boundaries in $K$: at high information ratios a soft classifier should observe longer than a hard classifier, and at low ratios it should stop sooner; the paper demonstrates this for the L1 and cross-entropy examples without proving a general ordering result.","The convex-envelope characterization suggests a computational route valid beyond the two-boundary case: compute the largest convex minorant of $H$ by a one-dimensional convex hull, then read off the value function and the contact set as the stopping region even if (G2) fails and the region disconnects.","If the information ratio is made endogenous or time-dependent, the constant thresholds would become moving boundaries and the phase transition at $K=\\beta^{-1}$ would become a separating surface in state-time space; the free-boundary formulation in this paper is the natural starting point for such an extension."],"forward_implications":["The optimal procedure is a two-threshold rule on the posterior: while $\\Pi_t$ lies in $(A^*,B^*)$ the observer keeps sampling, and the smallest optimal stopping time is the first exit from that interval.","There is a sharp phase transition: when the squared signal-to-noise ratio $K$ is at or below the threshold $\\beta^{-1}$, immediate stopping is optimal and no observation is worthwhile; for the L1 and cross-entropy penalties this threshold is $K=8$, whereas the hard-classification problem has a non-empty waiting region for every $K>0$.","As $K\\to\\infty$, the boundaries move to the endpoints: $A^*(K)\\downarrow0$ and $B^*(K)\\uparrow1$, with the rate bounds $A^*(K)\\le1/(1+CK^{1-\\epsilon})$ and $B^*(K)\\ge CK^{1-\\epsilon}/(1+CK^{1-\\epsilon})$; as $K$ falls to $\\beta^{-1}$, both boundaries collapse to the minimizer $\\pi_0$.","The value function can be written as $2K^{-1}\\Psi$ plus the convex envelope of $H$, so the whole optimal-stopping solution is encoded in a single convex-minorant computation; in symmetric penalties this reduces further to $B^*=1-A^*$."],"supporting_citations":[{"why":"Supplies the free-boundary formulation, posterior-process dynamics, and the verification template that Theorems 2.1 and 3.1 adapt.","marker":"[16]"},{"why":"Supplies the reduction to regular stopping times and the comparison argument for the immediate-stop case of Theorem 3.1.","marker":"[13]"},{"why":"Supplies the bound used to justify the martingale property of the stochastic integral in the verification step.","marker":"[2]"},{"why":"Supplies the common-tangent theorem for strictly separated convex sets used in Lemma 2.5 to prove uniqueness of the boundary pair.","marker":"[4]"},{"why":"Supplies the alternative common-tangent result invoked alongside [4] in the proof of Lemma 2.5.","marker":"[15]"}],"fun_headline_variants":["Exact solution found for Bayesian soft classification of Brownian drift","Two boundaries settle optimal stopping in drift classification","Soft classification: stopping rule fully characterized","When to stop classifying a Brownian drift? Exact answer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the single-well curvature condition (G2): the weighted second derivative $Ag(\\pi)=\\frac12\\pi^2(1-\\pi)^2g''(\\pi)$ must be strictly decreasing then strictly increasing with one minimum, because this is what forces the waiting region to be a single interval and makes the two-boundary solution unique.","fun_headline_variants_meta":{"raw":{"variants":["Exact solution found for Bayesian soft classification of Brownian drift","Two boundaries settle optimal stopping in drift classification","Soft classification: stopping rule fully characterized","When to stop classifying a Brownian drift? Exact answer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000498,"raw_usage":{"total_tokens":2407,"prompt_tokens":883,"completion_tokens":1524,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":1463}},"tokens_in":499,"tokens_out":1524,"duration_ms":10450,"temperature":1.0,"reasoning_tokens":1463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:25:40.888985+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete numerical check: for the L1 penalty $g(\\pi)=2\\pi(1-\\pi)$ with running cost $c=1$ and squared signal-to-noise ratio $K=7$, the theorem says $V^*\\equiv g$ and immediate stopping is optimal; a simulation-based approximation of (4) that finds any bounded stopping rule with expected cost below $g(\\pi_0)$ at some starting prior would refute the verification theorem. At $K=9$ the same simulation should reproduce the expected cost of first exit from $(A^*,B^*)$ obtained from (17)–(18).","supporting_citations":[{"cited_title":"Peskir and A","cited_arxiv_id":null,"evidence_quote":"Supplies the free-boundary formulation, posterior-process dynamics, and the verification template that Theorems 2.1 and 3.1 adapt."},{"cited_title":"Irle and V","cited_arxiv_id":null,"evidence_quote":"Supplies the reduction to regular stopping times and the comparison argument for the immediate-stop case of Theorem 3.1."},{"cited_title":"Common tangents to convex bodies","cited_arxiv_id":"2108.13569","evidence_quote":"Supplies the common-tangent theorem for strictly separated convex sets used in Lemma 2.5 to prove uniqueness of the boundary pair."},{"cited_title":"Lewis, B","cited_arxiv_id":null,"evidence_quote":"Supplies the alternative common-tangent result invoked alongside [4] in the proof of Lemma 2.5."}],"review_version":1}