{"id":"718e85c0-dd56-4cff-8c35-a3ab7adc26a3","arxiv_id":"1908.07626","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"For power-utility investors facing stocks driven by two correlated volatility factors, the value function and an O(epsilon)-accurate optimal policy are obtained from a regular perturbation around perfectly correlated factors, with a proven O(epsilon-squared) error bound.","lead":"This paper tackles a portfolio optimization problem where a stock's return and volatility are driven by two correlated random factors, and approximates the optimal value function by a first-order expansion around the case where the factors are perfectly correlated. The expansion reduces a nonlinear equation to two linear ones, and the authors prove an error bound of order epsilon-squared and provide a nearly optimal trading strategy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The order-epsilon residuals (31) and (33) display -f_1 while the defining PDE (17) has +f_1; with the printed sign the central cancellation in Theorem 3 does not occur.","rationale":"After checking the reader's stated concerns, the p<0 sign objection does not survive: because epsilon^2 M(-t) is negative, v_- has the larger base argument and, multiplied by x^p/p < 0, is indeed the smaller function; the terminal inequalities in Section 4.2.5 are consistent. The example-outside-assumptions point is acknowledged by the authors in Remark 2 and is a limitation rather than a hidden flaw in Theorem 3. The most load-bearing issue is different and internal: the displayed order-epsilon residual in the sub-super-solution computation has the wrong sign relative to equation (17). If the sign is corrected, the remaining gap is the omitted O(epsilon^3) estimate and martingale verification, which are standard but essential. Thus the reader's conditional verdict is the right one, but the paper as printed does not establish the advertised rigorous accuracy result.","tokens_in":17812,"tokens_out":20808,"duration_ms":179534,"concrete_test":"Use a symbolic algebra system to expand the HJB residual in (10) after inserting (16) (and, for the sub-solution, the Q^(pi^0) operator), retaining only the order-epsilon coefficient with correlations (15). Then substitute (17). If the coefficient is +(partial_t Psi^1 + L^(1,rho) Psi^1 + (Gamma/(2q)) lambda^2 Psi^1 + f_1) times q (x^p/p) (Psi^0)^(q-1), the printed -f_1 is a typo and the cancellation is restored; if it is -(partial_t Psi^1 + L^(1,rho) Psi^1 + (Gamma/(2q)) lambda^2 Psi^1 + f_1), the residual is -2 f_1 and the proof's central step fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central cancellation in Theorem 3 is printed with the wrong sign. Section 4.2.2/4.2.3 claims that the order-1 and order-epsilon terms in (31) and (33) cancel by (12) and (17). But (17) reads partial_t Psi^(1) + L^(1,rho) Psi^(1) + (Gamma/(2q)) lambda^2 Psi^(1) + f_1 = 0, while the brackets displayed in (31) and (33) are partial_t Psi^(1) + L^(1,rho) Psi^(1) + (Gamma/(2q)) lambda^2 Psi^(1) - f_1. Substituting (17) gives -2 f_1, not 0. This is an O(epsilon) residual, so no choice of M can absorb it for small epsilon; the O(epsilon^2) accuracy claim and near-optimality (30) do not follow as written. The same -f_1 appears in the multi-asset residual in Section 5.1.2. This may be a fixable sign typo, but it is an internal inconsistency in the proof of the paper's main theorem. The proof also still delegates the O(epsilon^3) sign control and the martingale verification to 'straightforward' omitted computations, so the rigorous accuracy advertised in the abstract is not fully delivered.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the Merton optimal investment problem with power utility in a market with one risky asset and two stochastic volatility factors. Since the multi-factor HJB equation is not tractable, the authors expand around the case of perfectly correlated factors. They perturb the correlations as rho_i = rho + epsilon rho_i^(1) and rho_12 = 1 + epsilon rho_12^(1), posit the ansatz v(t,x,z1,z2) ~ (x^p/p)(Psi^(0)+epsilon Psi^(1))^q, and derive linear PDEs (12) and (17) for Psi^(0) and Psi^(1). The paper gives a closed-form example in a Chacko-Viceira-type model and presents numerical error plots. The central theoretical claim is Theorem 3: under boundedness and regularity assumptions, v differs from the first-order approximation by x^p O(epsilon^2), uniformly in the state variables, and the proposed strategy pi^0 in (19) is nearly optimal in the sense of (30). An extension to two assets and two factors is sketched in Section 5.","tokens_in":17960,"tokens_out":5926,"duration_ms":242456,"significance":"If correct, the paper offers a tangible reduction in computational complexity: the nonlinear HJB equation is replaced by two linear equations, and an explicit nearly optimal strategy is identified. The sub- and super-solution framework is a natural and potentially reusable way to turn a heuristic asymptotic expansion into a uniform error bound. The main strengths are the explicit derivation of Psi^(0) and Psi^(1), the closed-form example, and the clear architecture of the accuracy proof. However, as printed, the proof of Theorem 3 contains a sign inconsistency in the order-epsilon cancellation, and it delegates the decisive O(epsilon^3) sign control and the martingale verification to omitted computations. These issues are local and likely fixable, but they prevent the rigorous accuracy claim from being fully supported in the current version.","major_comments":[{"comment":"The order-epsilon bracket in (31) is printed as partial_t Psi^(1) + L^{1,rho} Psi^(1) + (Gamma/(2q)) lambda^2 Psi^(1) - f_1(Psi^(0), nabla Psi^(0), H(Psi^(0))), and the same expression appears in (33). But the defining equation (17) has this expression with + f_1. Substituting (17) therefore gives -2 f_1, not 0, so the claimed cancellation of order-epsilon terms does not occur. This is not a harmless sign in an intermediate step: it leaves an O(epsilon) residual that cannot be absorbed by the M-dependent epsilon^2 term for small epsilon, and the proof of Theorem 3 and of (30) fails as written. The same mismatch occurs in the multi-asset formal computation in Section 5.1.2 following Eq. (42). One of the signs in (17), (31), (33), or in the definition of f_1 must be corrected, and the cancellation must be verified with the corrected sign.","section":"Section 4.2.2, Eq. (31); Section 4.2.3, Eq. (33)"},{"comment":"The proof relies on two claims that are explicitly not shown: that the O(epsilon^3) terms in (31) and (33) do not change the sign of the epsilon^2 term for sufficiently small epsilon, and that the martingale parts in the Ito arguments leading to (26) and (27) are true martingales. The manuscript says 'We omit here this lengthy but straightforward computation' and 'We omit the details.' These are load-bearing for the advertised rigorous accuracy result. The O(epsilon^3) remainder must be shown to be uniformly controlled in x, t, z, and M, and the martingale property must be checked using the stated boundedness and admissibility hypotheses. As printed, the theorem is not fully proven.","section":"Section 4.2.2 and Section 4.2.3, proof of Theorem 3"},{"comment":"The theorem's hypotheses exclude the paper's only fully worked example. The square-root processes have alpha_i(z)=m-z and beta_i(z)=beta-bar sqrt(2z), so beta_i is neither bounded nor bounded away from zero; hence Lemma 1 and Theorem 3 do not apply, as Remark 2 admits. The numerical O(epsilon) and O(epsilon^2) errors in Figures 1 and 2 are therefore not justified by the theorem. The authors should either extend the accuracy result to the example with the stopping argument they mention, or explicitly state that the numerical illustration is heuristic and outside the scope of the rigorous theorem. The abstract's unqualified claim of a rigorous accuracy result should be qualified accordingly.","section":"Section 4.2.4, Remark 2, and Section 3"}],"minor_comments":[{"comment":"There are several typographical errors, including 'pratical' in Section 3, 'begining' in Section 4.2, 'exemple' in Section 5.1.3, and 'there exits a constant M' in Theorem 3. These should be corrected in a final revision.","section":"Throughout"},{"comment":"The displayed computation after (43) contains a confusing fraction 'L^{pi0,rho^W,rho}_{x,z} v / v' where the division by v appears to be a typographical artifact; the equation should be presented without that notation.","section":"Section 5.1.2"},{"comment":"The text 'Psi^(0' is missing a closing brace; it should read 'Psi^(0)'.","section":"Remark 2"},{"comment":"The paper assumes rho_i = rho + rho_i^(1) epsilon for i=1,2 with the same base rho. It would be helpful to state explicitly that the base rho satisfies |rho|<1 and that the perturbed coefficients preserve the covariance condition (3) uniformly in epsilon, not merely for fixed epsilon.","section":"Section 2.2, Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"The sign inconsistency in Eqs. (31) and (33) is the most serious issue, but it appears to be a fixable typo rather than a fatal flaw in the underlying method. The omitted O(epsilon^3) and martingale estimates are also fixable, though they must actually be supplied before the paper can claim a rigorous accuracy result. I would not recommend rejection if the authors can complete the proof and clarify the scope of the theorem relative to the worked example."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The correlation-based perturbation is the genuinely new piece here. Expanding around fully correlated volatility factors rather than using time-scale separation is a real contribution, and it gives a tractable way to handle multiple same-timescale factors. The first-order correction Ψ^(1) solving a linear PDE is new, and the sub/super-solution approach is appropriate for the nonlinear HJB equation. If Theorem 3 is correct, the near-optimality of π0 is a genuine result.\n\nBut the proof as printed does not deliver. In Section 4.2.2, formula (31) has the order-ε bracket written with −f1, while Ψ^(1) is defined in (17) with +f1. Substituting (17) leaves an O(ε) residual; the claimed cancellation does not occur. The same −f1 appears in the multi-asset residual in Section 5.1.2. This looks like a fixable sign typo, but it is in the central step of the main theorem, so the accuracy result is unproven as written.\n\nThere are other soft spots, roughly in decreasing order of concern. The proof of Theorem 3 punts on the O(ε^3) sign estimate and on the martingale verification, both needed to justify the absorption by M. The paper's own Remark 2 concedes that the square-root example in Section 3 does not satisfy the theorem's assumptions, so the 'errors of order O(ε) and O(ε^2)' in Figures 1–2 are heuristic; the numerical method behind the reference solution is also undocumented. In Section 4.2.5 (p<0), the claim that v− < v+ appears to have the sign backwards: with p<0, x^p/p is negative, so the defined functions satisfy the opposite inequality. None of this is fatal to the method—the 0<p<1 bounded-coefficient version is plausibly correct once the sign is fixed and the omitted steps are supplied—but the manuscript does not currently support the advertised rigor.\n\nI'd send this to a serious referee, with instructions to check the sign and ask for the missing estimates. The idea is fresh and the intended audience—people doing asymptotic expansions in stochastic portfolio theory—will care. The current version should not be accepted as is.","headline":"Novel correlation-based perturbation with a promising accuracy theorem, but the proof as printed has a sign inconsistency in the central cancellation and several key verifications are omitted.","tokens_in":18642,"tokens_out":4904,"would_cite":false,"duration_ms":473911,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91G80","60H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that for a power-utility investor facing two correlated stochastic volatility factors, expanding around the perfectly correlated case yields an ε²-accurate approximation of the optimal value and a nearly optimal trading…","keywords":["optimal investment","stochastic volatility","power utility","distortion transformation","asymptotic expansion","HJB equation","near-optimal strategy","sub- and super-solutions"],"falsifier":"Choose a two-factor model whose coefficients satisfy Lemma 1 and Theorem 3, solve the HJB equation numerically to high accuracy, and compute $|v - (x^p/p)(\\Psi^{(0)}+\\varepsilon\\Psi^{(1)})^q|/x^p$ for $\\varepsilon = 0.1, 0.05, 0.025$; if this normalized error does not decay like a constant times $\\varepsilon^2$ uniformly in time and the factors, the theorem is false. Testing the square-root example instead would not settle the theorem, since the paper explicitly excludes it from the hypotheses.","tokens_in":17409,"feed_emoji":"📈","tokens_out":10757,"duration_ms":98848,"temperature":0.7,"pith_summary":"The paper treats a portfolio allocation problem in which a stock's return and volatility are driven by two correlated random factors, with the investor maximizing power utility of terminal wealth. When there is only one factor, a classical distortion transformation linearizes the Hamilton–Jacobi–Bellman equation, but with several factors the equation stays nonlinear. The paper's idea is to perturb around the limiting case where the two factors are perfectly correlated, because that limit reduces to the solvable single-factor problem and produces an explicit zeroth-order solution. The main result is a theorem saying that the value function is approximated to order ε² uniformly in wealth and factor values, and that the simple zeroth-order strategy is nearly optimal in the same sense. What makes this useful is that the approximation only requires solving two linear equations in lower dimension, not the fully nonlinear HJB equation.","feed_headline":"Two linear equations solve the multi-factor portfolio problem to ε²","feed_subtitle":"Instead of solving a fully nonlinear control problem, solve two linear equations and get a nearly optimal trading strategy.","key_machinery":"The load-bearing object is the distortion transformation $v(t,x,z_1,z_2) = (x^p/p)(\\Psi(t,z_1,z_2))^q$ with $q = 1/(1+\\Gamma\\rho^2)$ and $\\Gamma = p/(1-p)$, which converts the power-utility HJB equation into an equation for $\\Psi$. In the perfectly correlated limit $\\rho_1=\\rho_2=\\rho$, $\\rho_{12}=1$, the choice of $q$ cancels the nonlinear terms and $\\Psi^{(0)}$ solves a linear parabolic equation with a Feynman–Kac representation. The perturbation in $\\varepsilon$ uses the ansatz $\\Psi \\approx \\Psi^{(0)}+\\varepsilon\\Psi^{(1)}$; $\\Psi^{(1)}$ solves the same linear operator applied to a source $f_1$ built from $\\Psi^{(0)}$ and its derivatives. The accuracy proof then forms $v^\\pm = (x^p/p)(\\Psi^{(0)}+\\varepsilon\\Psi^{(1)}\\pm\\varepsilon^2 M(T-t))^q$ and shows that $v^-$ is a submartingale along $\\pi^0$ while $v^+$ is a supermartingale along every admissible strategy, which sandwiches the true value function within $x^pO(\\varepsilon^2)$.","core_discovery":"On the paper's own terms, the central result is Theorem 3. Under boundedness, smoothness, and non-degeneracy assumptions on the coefficients, with the return-to-volatility ratio $\\lambda$ bounded away from zero and $0<p<1$, the value function satisfies $v(t,x,z_1,z_2) = (x^p/p)(\\Psi^{(0)}(t,z_1,z_2)+\\varepsilon\\Psi^{(1)}(t,z_1,z_2))^q + x^pO(\\varepsilon^2)$, uniformly in time and factors. The same theorem asserts that the strategy $\\pi^0$ built from the zeroth-order solution is nearly optimal: $0 \\le v(t,x,z_1,z_2) - E_{t,x,z_1,z_2}[(1/p)(X_T^{\\pi^0})^p] = x^pO(\\varepsilon^2)$. The function $\\Psi^{(0)}$ is the solution of a linear equation inherited from the perfectly correlated limit, $\\Psi^{(1)}$ solves a companion linear equation whose source term is built from $\\Psi^{(0)}$ and its derivatives, and the proof works by constructing sub- and super-solutions whose difference is of order $\\varepsilon^2$. The same expansion is carried out formally for two stocks and two factors.","pith_inferences":["If the stopping-and-truncation argument mentioned in Remark 2 can be made uniform in the truncation parameter, the ε² accuracy should extend to square-root volatility factors, bringing the paper's explicit example under the theorem.","A likely extension, not pursued in the paper, is to apply the same perfect-correlation perturbation to other settings that admit a distortion transformation, such as long-run or infinite-horizon problems, where the limiting one-factor model is still linear.","A practical consequence the authors do not state: the error bound's constant depends on how far the Sharpe ratio and factor volatilities stay away from zero, so the approximation should be trusted least in volatility states near zero; a finite-horizon stopping version could quantify that degradation."],"forward_implications":["The optimal value function in a multi-factor power-utility problem can be computed to order ε² by solving two linear parabolic equations, one for Ψ⁽⁰⁾ and one for Ψ⁽¹⁾, rather than the fully nonlinear HJB equation.","The zeroth-order strategy π⁰ delivers expected terminal utility within O(ε²) of the true optimum, so an investor can act on a simple explicit rule without solving the nonlinear control problem.","The approximation error is uniform over the factor domain covered by the hypotheses, so the method is not limited to a particular calibrated point.","The same ansatz and linear-equation structure extend to two stocks and two factors, indicating the approach is not tied to a single risky asset."],"supporting_citations":[{"why":"Supplies the distortion-transformation linearization and the zeroth-order expansion that the paper adapts to the perfectly correlated limit.","marker":"[13]"},{"why":"Introduces the distortion transformation used to convert the power-utility HJB equation into a linear equation for the single-factor case.","marker":"[20]"},{"why":"Provides the stochastic-volatility consumption/portfolio model whose explicit square-root extension generates the worked example and the formulas for Ψ⁰ and Ψ¹ on the diagonal.","marker":"[4]"},{"why":"Gives the Feynman–Kac representations and boundedness results used in Lemma 1 to control Ψ⁰, Ψ¹, and their derivatives.","marker":"[14]"},{"why":"Cited for existence and uniqueness of classical solutions of degenerate elliptic/parabolic equations, the regularity input for Lemma 1.","marker":"[18]"},{"why":"Provides the asymptotic sub- and super-solution verification technique that the accuracy proof in Section 4 adapts.","marker":"[2]"},{"why":"Extends the sub- and super-solution construction to finite-horizon optimal investment with stochastic volatility, the direct template for Theorem 3's proof.","marker":"[3]"}],"fun_headline_variants":["Two linear equations beat nonlinear HJB to ε²","Multi-factor portfolio: solve two linear PDEs, get ε² accuracy","Correlated factors? Two linear equations suffice for ε² value","Perturbation trick slashes nonlinear control to two linear equations","Approximate optimal investment with two linear solves, error ε²"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theorem assumes the Sharpe ratio, the factor drifts, and the factor volatilities are bounded with bounded derivatives, with the volatilities and the Sharpe ratio bounded away from zero; the entire sub- and super-solution argument and the Feynman–Kac representations depend on those bounds, and the paper's own square-root example does not satisfy them.","fun_headline_variants_meta":{"raw":{"variants":["Two linear equations beat nonlinear HJB to ε²","Multi-factor portfolio: solve two linear PDEs, get ε² accuracy","Correlated factors? Two linear equations suffice for ε² value","Perturbation trick slashes nonlinear control to two linear equations","Approximate optimal investment with two linear solves, error ε²"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000992,"raw_usage":{"total_tokens":4224,"prompt_tokens":984,"completion_tokens":3240,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":3153}},"tokens_in":600,"tokens_out":3240,"duration_ms":23636,"temperature":1.0,"reasoning_tokens":3153,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:02:19.300749+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a two-factor model whose coefficients satisfy Lemma 1 and Theorem 3, solve the HJB equation numerically to high accuracy, and compute $|v - (x^p/p)(\\Psi^{(0)}+\\varepsilon\\Psi^{(1)})^q|/x^p$ for $\\varepsilon = 0.1, 0.05, 0.025$; if this normalized error does not decay like a constant times $\\varepsilon^2$ uniformly in time and the factors, the theorem is false. Testing the square-root example instead would not settle the theorem, since the paper explicitly excludes it from the hypotheses.","supporting_citations":[{"cited_title":"Fouque, R","cited_arxiv_id":null,"evidence_quote":"Supplies the distortion-transformation linearization and the zeroth-order expansion that the paper adapts to the perfectly correlated limit."},{"cited_title":"Zariphopoulou","cited_arxiv_id":null,"evidence_quote":"Introduces the distortion transformation used to convert the power-utility HJB equation into a linear equation for the single-factor case."},{"cited_title":"Chacko and L","cited_arxiv_id":null,"evidence_quote":"Provides the stochastic-volatility consumption/portfolio model whose explicit square-root extension generates the worked example and the formulas for Ψ⁰ and Ψ¹ on the diagonal."},{"cited_title":"Karatzas and S","cited_arxiv_id":null,"evidence_quote":"Gives the Feynman–Kac representations and boundedness results used in Lemma 1 to control Ψ⁰, Ψ¹, and their derivatives."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited for existence and uniqueness of classical solutions of degenerate elliptic/parabolic equations, the regularity input for Lemma 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the asymptotic sub- and super-solution verification technique that the accuracy proof in Section 4 adapts."},{"cited_title":"Bichuch and R","cited_arxiv_id":null,"evidence_quote":"Extends the sub- and super-solution construction to finite-horizon optimal investment with stochastic volatility, the direct template for Theorem 3's proof."}],"review_version":1}