Pith. sign in

REVIEW 2 major objections 4 minor 28 references

S-shaped Utility Maximization with VaR Constraint and Partial Information

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper identifies one computable critical wealth level that decides when a VaR-constrained S-shaped investor with unobservable drift has a unique optimal portfolio, a limiting claim, or no feasible strategy.

desk verdict Solid feasibility characterization for a two-state hidden-drift model, but Theorem 3.1's existence claim omits an admissibility check that is likely to fail. read the letter →

arxiv 2506.10103 v1 pith:BK6U7GPM submitted 2025-06-11 q-fin.MF math.OC

Dongmei Zhu , Ashley Davey , Harry Zheng This is my paper
classification q-fin.MFmath.OC MSC 93E2093E1191G8090C4649M29
keywords S-shapedutilityVaRconstraintpartialinformationBayesianfilterdualcontrolconcavificationLagrangemultipliercriticalwealthlevel
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper studies an investor with S-shaped (prospect-theory) utility who faces a value-at-risk constraint on terminal wealth and cannot observe the drift of the risky asset, only the price path. It shows that, when the unobservable drift takes one of two possible values with a known prior, the entire problem is governed by a single critical wealth level. Above that level the constrained problem has a unique optimal terminal wealth and a unique Lagrange multiplier; at the level the only solution is the limiting claim that pays the lower bound exactly when the economic state is bad; below it no admissible strategy satisfies the VaR constraint. A constructive proof of this threshold yields an exact numerical algorithm, and two approximation algorithms (Monte Carlo simulation and a neural-network PDE solver) are compared in examples. If the two-state assumption is dropped, the paper states the change-of-measure argument no longer works, so the threshold formula is specific to that model.

What carries the argument

The load-bearing device is a change of measure from the filtered probability measure $P$ to $Q$ under which the filtered drift's odds ratio $\Phi(t)=(\hat{\mu}(t)-\mu_l)/(\mu_h-\hat{\mu}(t))$ becomes a geometric Brownian motion, reducing the two-dimensional joint distribution of the dual state and the filtered drift to one dimension. The Radon–Nikodym derivative $F(t)=(1+\Phi(t))/(1+\varphi)$ and the dual process $H(T)$ in (3.8) convert the dual value and constraint functions into the semi-closed integrals (3.4)–(3.5), from which $H^\*_\varepsilon$ and $\hat{x}_\varepsilon$ are read off.

What would settle it

Run a numerical search over admissible trading strategies for a parameter set with initial wealth strictly below the paper's $\hat{x}_\varepsilon$ (for instance the Section 5 parameters with $x_0=0.6$, $\varepsilon=0.2$, where $\hat{x}_\varepsilon\approx0.66$); any strategy achieving $P(X(T)\geq L)\geq1-\varepsilon$ would disprove the infeasibility claim. For the solvable direction, simulate $H(T)$ and $F(T)$ under $Q$ from (3.8) and check that $\mathbb{E}^Q[F(T)\mathbf{1}_{\{H(T)\leq H^\*_\varepsilon\}}]=1-\varepsilon$ and that the constructed terminal wealth satisfies the budget constraint.

Watch

Extended reading notes

Core claim

The central discovery is Theorem 3.1, which characterizes the VaR-constrained S-shaped utility problem under partial information by a single critical wealth level. Let $H^\*_\varepsilon$ solve $\mathbb{E}^Q[F(T)\mathbf{1}_{\{H(T)\leq H^\*_\varepsilon\}}]=1-\varepsilon$ and set $\hat{x}_\varepsilon := \mathbb{E}^Q[F(T)L\mathbf{1}_{\{H(T)<H^\*_\varepsilon\}}(1+\varphi)H(T)]$. Then initial wealth $x_0>\hat{x}_\varepsilon$ yields a unique optimal terminal wealth $X^{\pi^\*,\lambda^\*}(T)=x^\*_{\lambda^\*}(Y(T))$ with a unique Lagrange multiplier $\lambda^\*\geq0$; $x_0=\hat{x}_\varepsilon$ leaves only the limiting claim $L\mathbf{1}_{\{H(T)<H^\*_\varepsilon\}}$ almost surely; and $x_0<\hat{x}_\varepsilon$ makes the problem infeasible. The proof is constructive, splitting into three regions of $H^\*_\varepsilon$ relative to the concavified utility's breakpoints and producing the multiplier formulas (3.16) and (3.17).

Load-bearing premise

The whole threshold formula assumes the unobservable drift has exactly two possible values with known probabilities, independent of the price shocks; if the drift follows any richer distribution, the paper's key change of measure no longer works.

Editorial extensions

If this is right

  • The numerical algorithms give a practical way to compute optimal terminal wealth for S-shaped investors with VaR constraints when only price data are available.
  • The threshold $\hat{x}_\varepsilon$ can serve as a capital-requirement diagnostic: initial wealth below it cannot satisfy the VaR constraint no matter which admissible strategy is chosen.
  • The explicit optimal terminal wealth form shows the investor holds a quantile-like claim: a constant payoff $L$ in the bad state plus a concave utility payoff in better states, directly generalizing the fully observable solutions of Dong and Zheng (2020).
  • The result extends the concavification-plus-dual method to partial information for a two-state prior, giving an exact baseline against which approximate methods can be tested.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: The same odds-ratio change-of-measure idea should extend to Markov-modulated drift with a finite state space via the Wonham filter, although the threshold would lose its closed form and become a numerical object.
  • Editorial inference: The critical wealth level $\hat{x}_\varepsilon$ could be interpreted as a minimal capital requirement for a prospect-theory investor under regulatory VaR, a quantity one could estimate from historical price data and compare with observed minimum entry wealth.
  • Editorial inference: The dual PINN method trains the value function with the Lagrange multiplier as an input parameter, a design that could be reused for other constrained control problems whose dual PDE is linear.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper studies a continuous-time expected-utility maximization problem with an S-shaped (reference-dependent) utility, a Value-at-Risk constraint P(X(T)≥L)≥1−ε, and an unobservable drift that takes two values with a known prior. Using the Bayesian filter, the authors transform the model into a fully observable one with a filtered drift state, then apply the concavification principle and dual methods. A change of measure under which the odds ratio of the two filter states is a geometric Brownian motion reduces the dual problem to one dimension and yields semi-closed integral representations for the dual value and constraint functions, a critical wealth level x̂_ε separating feasibility from infeasibility, and a constructive existence/uniqueness theorem for the optimal terminal wealth and Lagrange multiplier. Three numerical algorithms are proposed and compared: an exact Lagrange algorithm, a dual Monte Carlo method, and a physics-informed neural network (PINN) method.

Significance. If the main results are sound, the paper gives a fully characterized, computationally implementable solution to a VaR-constrained S-shaped utility problem under partial information, extending Dong and Zheng (2020) to a non-observable drift. The paper's strengths include a constructive proof of Theorem 3.1 with explicit expressions for the budget function and the multiplier, a clearly stated infeasibility threshold, and reproducible numerical experiments with publicly available code. The Bernoulli-prior restriction is explicitly acknowledged in Section 6. However, two load-bearing issues remain: the asserted optimal control is not shown to be admissible, and the integral representation for t>0 is not correct as written. These issues are local and, in my view, fixable, but they prevent acceptance in the current form.

major comments (2)
  1. [Section 3, Theorem 3.1; Section 2, Eq. (2.1)] The proof of Theorem 3.1 never verifies that the control π*,λ* obtained from the martingale representation theorem belongs to the admissible class A defined by E∫_0^T |π(t)|² dt < ∞ in (2.1). The optimal terminal wealth in (3.14)–(3.15) contains cash-or-nothing discontinuities at the thresholds H(T)=c_z/(y0(1+φ)) and H(T)=k_λ/(y0(1+φ)), and the replicating portfolio proportion for such a digital claim behaves like (T-t)^{-1} on a set of paths of positive probability as t→T. Square-integrability of the dollar integrand does not imply square-integrability of the proportion when the wealth process can approach zero, and no bound on E∫|π*|² dt is provided. Thus the assertion that problem (2.6) has an optimal solution in A is not established. The authors should either prove the square-integrability of the candidate π* (possibly under additional parameter conditions) or reformulate the admissibility condition and state the theorem accordingly.
  2. [Section 3, Eqs. (3.4)–(3.5)] The function Ψ in (3.4)–(3.5) is defined as Ψ(t,x,μ̂)=(1+φ exp{Θx−½Θ²(T−t)})/(1+φ) with φ=Φ(0), so the right-hand sides of (3.4)–(3.5) do not depend on the current filtered estimate μ̂(t). The left-hand side v_c^λ(t,y,μ̂) in (2.24) is a conditional expectation given μ̂(t), and for a nonlinear dual function V_c^λ this expectation genuinely depends on μ̂(t). Consequently the stated integral representation cannot hold for general t<T; it is only valid at t=0, where Φ(0)=φ. The correct time-t formula should replace φ by Φ(t)=(μ̂(t)−μ_l)/(μ_h−μ̂(t)) in the definition of Ψ, i.e., Ψ_t(x)=(1+Φ(t)e^{Θx−½Θ²(T−t)})/(1+Φ(t)). As printed, the advertised 'semi-closed integral representation for the dual value function' is inaccurate for t>0, and (3.5) has the same defect.
minor comments (4)
  1. [Section 5.1] The data paragraph lists λ=0.2 as an input parameter, but λ is the Lagrange multiplier that should be determined by the quantile constraint; this is confusing and should be clarified or removed.
  2. [Section 3, Eqs. (3.4)–(3.5)] The displayed formula in plain text is ambiguous: the argument of V_c^λ should be the fraction y exp{...}/Ψ(t,x,μ̂), not a product. Please fix the typesetting so that the division is clear.
  3. [Section 4.3, Eq. (4.21)] The notation |σ^{-1}(μ−r)|² is nonstandard; write (σ^{-1}(μ−r))².
  4. [Section 3, after Eq. (3.7)] The sentence 'Y(T)=Y(T)/F(T)=y0(1+φ)H(T)' uses Y(T) with two different meanings (the P-dual process on the left and the Q-process on the right); please clarify the notation.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the critical wealth threshold and Lagrange multiplier are solved from explicit model equations, not fitted; self-citations are not load-bearing.

full rationale

After walking the derivation chain, I find no circular step that reduces a prediction to an input. The main theorem's threshold x̂_ε is an explicit Q-expectation of the digital claim L1_{H<H*_ε} times (1+φ)H(T), and H*_ε is defined by the quantile equation (3.10); the proof solves for y0 from the budget constraint (3.9) and for λ* from the binding quantile condition, using monotonicity of the relevant functions. No parameter is fitted to data and then renamed a prediction. The two imported blocks, Proposition 2.1 (Dong and Zheng 2020) and the measure change (Xing et al. 2025), are co-authored by Zheng, but both are parameter-free published results with stated assumptions that do not include the present target result; the paper re-derives the Girsanov change of measure explicitly rather than merely citing it. These self-citations are therefore not load-bearing circularity. The unproved admissibility of the martingale-representation control π* and the restriction to a two-state prior are genuine correctness/scope concerns, but they are not circularity. Score 2 reflects the presence of minor self-citations, not a circular derivation.

Assumptions & free parameters 2 free parameters · 7 assumptions · 0 invented entities

The theorem is parameter-free: H*_ε, x̂_ε, λ* and y0 are all computed from explicit expectations under Q given the model inputs, nothing is fitted. The two free parameters listed are example settings for Section 5 only and do not enter the proof. The axioms are the standard probabilistic tools plus the domain assumptions: a Bernoulli drift independent of W, the filter SDE (2.5), the concavification equivalence, utility regularity, and L < θ. No invented entities appear: Q, Φ, F and H are auxiliary processes derived from the model, not new primitives.

free parameters (2)
  • Example market and utility parameters: U1(x)=√x, U2(x)=x^0.3, θ=1.5, L=0.9, x0=1.0, r=0.05, σ=0.2, T=1, μ̂(0)=0.07 = Hand-chosen for Section 5.1
    These instantiate the model in the numerical section; they are not fitted to data and do not appear in the proof of Theorem 3.1.
  • Algorithm hyperparameters: M=100000, N=100, δ=0.1/0.01, 10 or 100 hidden nodes, 2000+200 PINN points, ADAM, loss… = Given in Sections 4.2-5.1
    Hand-chosen settings for the numerical comparison in Section 5; they do not enter the mathematical claims.
assumptions (7)
  • standard math Girsanov theorem and the Q-martingale property of the density process F(t) = dP/dQ on F^S_t
    Invoked in Section 3 to define W^Q and to transform the dynamics of X, Φ, F and Y; standard and unproblematic.
  • domain assumption Two-state Bayesian filtering SDE (2.5): dμ̂(t) = σ^{-1}(μ̂−μ_l)(μ_h−μ̂)dŴ, with μ̂(t) ∈ (μ_l, μ_h) a.s.
    Section 2.1, cited to Sass (2007) and Décamps et al. (2005). The entire change-of-measure reduction depends on this filter equation and on the bounds of μ̂.
  • domain assumption The drift μ is F_0-measurable, Bernoulli on {μ_h, μ_l} with P(μ = μ_h) = p, and independent of the Brownian motion W
    Section 2, model paragraph. This two-state structure is the load-bearing premise; the authors concede in Section 6 that general priors break the method.
  • domain assumption Concavification equivalence: the non-concave unconstrained problem (2.7) has the same value and optimizer as the concavified problem (2.18)
    Section 2.2, applied via Reichlin (2013), Theorem 5.1; the paper assumes the utility satisfies the conditions of that theorem.
  • domain assumption Regularity of U1, U2: strictly increasing, strictly concave, C1, U_i'(0+)=∞, asymptotic elasticity lim_{x→∞} xU1'(x)/U1(x) < 1
    Section 2; used in Theorem 3.1's proof for the monotonicity of the budget function and the finiteness of the dual expectations.
  • standard math Feynman-Kac formula identifying the dual value function (2.24) as the solution of the linear PDE (2.26)
    Section 2.2; used by the PINN algorithm in Section 4.3.
  • domain assumption L < θ, the VaR floor lies strictly below the reference point
    Section 2; the case L ≥ θ is deferred and not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of S-shaped Utility Maximization with VaR Constraint and Partial Information." pith.science (2026). https://pith.science/paper/BK6U7GPM

@misc{pith2026250610103,
  author       = {Pith},
  title        = {Pith review of: S-shaped Utility Maximization with VaR Constraint and Partial Information},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BK6U7GPM}},
  note         = {Machine review of arXiv:2506.10103}
}
read the original abstract

We study S-shaped utility maximisation with VaR constraint and unobservable drift coefficient. Using the Bayesian filter, the concavification principle, and the change of measure, we give a semi-closed integral representation for the dual value function and find a critical wealth level that determines if the constrained problem admits a unique optimal solution and Lagrange multiplier or is infeasible. We also propose three algorithms (Lagrange, simulation, deep neural network) to solve the problem and compare their performances with numerical examples.

Figures

Figures reproduced from arXiv: 2506.10103 by the authors.

Figure 1
Figure 1. Utility Uλ and concavified utility U c λ : (a) kλ ≤ U ′ 1 (˜z − θ) and (b) kλ > U′ 1 (˜z − θ). Now we consider the auxiliary stochastic control problem with fixed λ:    supπ∈A E[U c λ (Xπ (T))], s.t. Xπ (t) satisfies (2.4), µˆ(t) satisfies (2.5). (2.18) Denote the value function of (2.18) by u c λ (t, x, µˆ) . .= sup π∈A E[U c λ (Xπ (T))|Xπ (t) = x, µˆ(t) = ˆµ] (2.19) and the constraint probability function h by … view at source ↗
Figure 2
Figure 2. Numerical results for the concavified problem. [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Distribution of Xπ ∗,λ(T) for three different values of λ, log scaled. 5.3 Problem feasibility For fixed ε = 0.2, we may compute the value of ˆxε defined in Theorem 3.1 as ˆxε ≈ 0.66. This is the minimum wealth needed to achieve P(Xπ ∗,λ(T) ≥ L) = 0.8 for some λ ≥ 0 [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Results for dual simulation and discrete algorithm in both feasible and infeasible regions. [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 27 canonical work pages

  1. [1]

    and Crisan, D

    Bain, A. and Crisan, D. (2009). Fundamentals of Stochastic Filtering, Springer, New York

    work page 2009

  2. [2]

    Basak, S., Shapiro, A. (2001). Value-at-risk-based risk management: optimal policies and asset prices. The review of financial studies, 14(2), 371-405

    work page 2001

  3. [3]

    Bensoussan, A., Hoe, S., Kim, J., Yan, Z. (2022). A risk extended version of merton’s optimal consumption and portfolio selection. Operations Research, 70(2), 815-829

    work page 2022

  4. [4]

    B., Kouwenberg, R., Post, T

    Berkelaar, A. B., Kouwenberg, R., Post, T. (2004). Optimal portfolio choice under loss aversion. Review of Economics and Statistics, 86(4), 973-987

    work page 2004

  5. [5]

    Economic Dynamics Control 51, 28-49, 2015

    B Bian, H Zheng, Turnpike property and convergence rate for an investment model with general utility functions, J. Economic Dynamics Control 51, 28-49, 2015

    work page 2015

  6. [6]

    Brendle, S. (2006). Portfolio selection under incomplete information. Stochastic processes and their Applications, 116(5), 701-723

    work page 2006

  7. [7]

    Carpenter, J. N. (2000). Does option compensation increase managerial risk appetite?. The journal of finance, 55(5), 2311-2331

    work page 2000

  8. [8]

    Chen, A., Nguyen, T., Stadje, M. (2018). Risk management with multiple VaR constraints. Mathematical Methods of Operations Research, 88, 297-337

    work page 2018

Show all 28 references
  1. [9]

    Davey, A., Zheng, H. (2022). Deep learning for constrained utility maximisation. Methodology and Computing in Applied Probability, 24(2), 661-692

    2022

  2. [10]

    and Villeneuve, S

    D \'e camps, J.P., Mariotti, T. and Villeneuve, S. (2005). Investment timing under incomplete information. Mathematics of Operations Research, 30, 472--500

    2005

  3. [11]

    De Franco, C., Nicolle, J., Pham, H. (2019). Bayesian learning for the Markowitz portfolio selection problem. International Journal of Theoretical and Applied Finance, 22(07), 1950037

    2019

  4. [12]

    Detemple, J. B. (1986). Asset pricing in a production economy with incomplete information. The Journal of Finance, 41(2), 383-391

    1986

  5. [13]

    Dong, Y., Zheng, H. (2020). Optimal investment with S-shaped utility and trading and Value at Risk constraints: An application to defined contribution pension plan. European Journal of Operational Research, 281(2), 341-356

    2020

  6. [14]

    Ekstr\"om, E., Vaicenavicius, J. (2016). Optimal liquidation of an asset under drift uncertainty. SIAM Journal on Financial Mathematics, 7(1), 357-381

    2016

  7. [15]

    Han, J., Jentzen, A., E, W. (2018). Solving high-dimensional partial differential equations using deep learning. Proceedings of the National Academy of Sciences, 115(34), 8505-8510

    2018

  8. [16]

    E., Jin, L

    Ingersoll, J. E., Jin, L. J. (2013). Realization utility with reference-dependent preferences. The Review of Financial Studies, 26(3), 723-767

    2013

  9. [17]

    Behavioral portfolio selection in continuous time

    Jin, H., Zhou, X.Y.(2008). Behavioral portfolio selection in continuous time. Mathematical Finance, 18(3), 385-426

    2008

  10. [18]

    Kahneman, D., Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 363-391

    1979

  11. [19]

    Karatzas, I., Xue, X. X. (1991). A Note On Utility Maximization Under Partial Observations 1. Mathematical Finance, 1(2), 57-70

    1991

  12. [20]

    Pham, H. (2009). Continuous-time stochastic control and optimization with financial applications (Vol. 61): Springer Science& Business Media

    2009

  13. [21]

    ,Karniadakis G

    Raissi M.,Perdikaris P. ,Karniadakis G. E. . (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational physics, 378, 686-707

    2019

  14. [22]

    Reichlin, C. (2013). Utility maximization with a given pricing measure when the utility is not necessarily concave. Mathematics and Financial Economics, 7(4), 531-556

    2013

  15. [23]

    Rieder, U., Bäuerle, N. (2005). Portfolio optimization with unobservable Markov-modulated drift process. Journal of Applied Probability, 42(2), 362-378

    2005

  16. [24]

    Sass, J. (2007). Utility maximization with convex constraints and partial information. Acta Applicandae Mathematicae, 97(1-3), 221-238

    2007

  17. [25]

    ,Karniadakis G

    Shin Y., Darbon J. ,Karniadakis G. E. (2020). On the convergence of physics informed neural networks for linear second-order elliptic and parabolic type pdes. arXiv preprint arXiv:2004.01806

    2020 arXiv

  18. [26]

    Wang, C., Li, S., He, D., Wang, L. (2022). Is L^ 2 Physics informed loss always suitable for training physics informed neural network?. Advances in Neural Information Processing Systems, 35, 8278-8290

    2022

  19. [27]

    Xing, J., Ma, J., Zheng, H. (2025). A simple integral equation approach for optimal investment stopping problems with partial information, Mathematics of Operations Research, published online

    2025

  20. [28]

    Yiu, K. F. C. (2004). Optimal portfolios under a value-at-risk constraint. Journal of Economic Dynamics and Control, 28(7), 1317-1334

    2004

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.