← All quantitative finance open problems

Stochastic control / Construction problem

Provably efficient solvers for high-dimensional fully nonlinear HJB equations

Would make high-dimensional portfolio allocation, optimal execution and hedging computations reliable through explicit accuracy and resource guarantees.

OpenEditorially scopedFramework needs definitionCompiled 2026-11-09

Seed catalog · Showing the bundled record.

Problem statement

Identify broad, explicitly structured families of controlled diffusions whose fully nonlinear HJB equation ∂tu+sup⁡a∈A{ba⋅∇u+12tr⁡(σa(σa)⊤D2u)+fa}=0\partial_t u+\sup_{a\in A}\{b^a\cdot\nabla u+\tfrac12\operatorname{tr}(\sigma^a(\sigma^a)^\top D^2u)+f^a\}=0, with terminal condition u(T,⋅)=gu(T,\cdot)=g, admits an implementable solver with proved error and computational cost polynomial in state dimension and inverse accuracy. Include control-dependent diffusion and quantify the cost of optimization or training, rather than only proving that a small approximating network exists.

Assumptions and scope

  • Specify coefficient/payoff representations, their dimension-dependent regularity constants, control-set complexity, horizon and access to simulation or coefficient oracles.
  • Fix an error norm and domain or sampling distribution, and include the success probability for randomized algorithms.
  • Assume comparison and well-posedness for the chosen viscosity-solution class. Restrict input structure enough to avoid hiding exponential complexity in the data.
  • Separate representation, discretization, sampling and optimization errors; a numerical benchmark or an infinite-width limit is not a finite-runtime guarantee.

What would settle it?

An end-to-end polynomial-complexity theorem for a financially meaningful fully nonlinear class beyond existing restricted results, or a complexity lower bound under an explicit input and algorithm model locating the boundary of tractability.

Selected for a theoretical or computational target. A successful calibration or numerical benchmark alone does not establish the stated mathematical guarantee.

Status and evidence

Convergent probabilistic methods and dimension-independent approximation results exist under structural assumptions. Grohs–Herrmann study a restricted control-affine, quadratic-cost class. Hutzenthaler–Nguyen's 2025 result shows a curse of dimensionality for full-history recursive multilevel Picard approximations on one HJB example; it is not an impossibility theorem for all solvers. The unresolved target is efficient end-to-end computation for further specified fully nonlinear classes.

Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.

Sources and further reading

  1. Fahim, Touzi & Warin: A probabilistic numerical method for fully nonlinear parabolic PDEs ↗background

    Foundational probabilistic approximation framework for fully nonlinear equations; convergence is distinct from dimension-polynomial cost.

  2. Grohs & Herrmann: Deep neural network approximation for high-dimensional parabolic Hamilton–Jacobi–Bellman equations ↗partial progress · source checked 2026-11-09

    Scholarly abstract: affine control dependence, quadratic costs and bounded controls.

    Approximation/representation result for a restricted class; not an unrestricted training-time guarantee.

  3. Hutzenthaler & Nguyen: Full history recursive multilevel Picard approximations suffer from the curse of dimensionality for the Hamilton–Jacobi–Bellman equation of a stochastic control problem ↗partial progress · source checked 2026-11-09

    arXiv abstract: lower bound for the named MLP approximation method on an example equation.

    The obstruction is method- and problem-specific.