Paper: arXiv 2311.12239
Authors: Chuhao Sun, Asaf Cohen, James Stokes, Shravan Veerapaneni
Abstract
Neural networks are increasingly recognized as a powerful numerical solution technique for partial differential equations (PDEs) arising in diverse scientific computing domains, including quantum many-body physics. In the context of time-dependent PDEs, the dominant paradigm involves casting the approximate solution in terms of stochastic minimization of an objective function given by the norm of the PDE residual, viewed as a function of the neural network parameters. Recently, advancements have been made in the direction of an alternative approach which shares aspects of nonlinearly parametrized Galerkin methods and variational quantum Monte Carlo, especially for high-dimensional, time-dependent PDEs that extend beyond the usual scope of quantum physics. This paper is inspired by the potential of solving Hamilton-Jacobi-Bellman (HJB) PDEs using Neural Galerkin methods and commences the exploration of nonlinearly parametrized trial functions for which the evolution equations are analytically tractable. As a precursor to the Neural Galerkin scheme, we present trial functions with evolution equations that admit closed-form solutions, focusing on time-dependent HJB equations relevant to finance.
Complexity vs Empirical Score
- Math Complexity: 9.0/10
- Empirical Rigor: 2.0/10
- Quadrant: Lab Rats — high math complexity, limited empirical testing
Why this score: The paper is highly mathematically dense, featuring advanced concepts like Galerkin methods, high-dimensional PDEs, and quantum mechanics analogies, but focuses on theoretical derivations and closed-form solutions for trial functions with no reported backtests or implementation metrics.
Research Flowchart
flowchart TD Start["Research Goal: Solve high-dimensional HJB PDEs using<br>Neural Galerkin methods"] --> Method["Key Methodology:<br>Nonlinearly parametrized trial functions"] Method --> Inputs["Data/Inputs:<br>Time-dependent HJB equations<br>Finance applications"] Inputs --> Comp["Computational Process:<br>Analytically tractable evolution equations<br>Closed-form solutions"] Comp --> Outcomes["Key Findings:<br>1. Novel Neural Galerkin framework<br>2. Closed-form trial functions<br>3. Extension to high-dimensional PDEs"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper introduces a family of nonlinearly parametrized trial functions with analytically tractable evolution equations as a precursor to Neural Galerkin methods for solving high‑dimensional time‑dependent HJB equations in finance.
Source sentence
This paper is inspired by the potential of solving Hamilton-Jacobi-Bellman (HJB) PDEs using Neural Galerkin methods and commences the exploration of nonlinearly parametrized trial functions for which the evolution equations are analytically tractable. As a precursor to the Neural Galerkin scheme, we present trial functions with evolution equations that admit closed-form solutions, focusing on time-dependent HJB
Assumes 3
- The HJB equation admits a unique solution under standard regularity conditions on the data of the control problem. model
Source sentence
Under some conditions on the data of the problem, the value function V is the unique solution of the following PDE, called the HJB equation, [6] ( (∂tϕ)(t, x) + H(t, x, (Dxϕ)(t, x), (D2 xϕ)(t, x)) = 0, (t, x) ∈[0, T] × Rd, ϕ(T, x) = g(x), where Dxϕ is the gradient of ϕ and D2 xϕ is the Hessian matrix; the function H is the Hamiltonian associated with the problem, where in the case described above, is given by: H(t, x
- Asset dynamics follow a constant‑coefficient Black‑Scholes model with constant drift, volatility, and Sharpe ratio. market
Source sentence
Consider the dynamics where both S and Y follow constant coefficient Black-Scholes model: dSt = (r + λσ)Stdt + σStdBt, S0 = s0, dYt = b ⊙Ytdt + a ⊙YtdWt, Y0 = y0 > 0, where λ = µ −r σ . is the Sharpe ratio for the risky asset S, and µ, σ are constants.
- The trial functions are sufficiently regular (L2 integrable) so that the integrals defining M and V converge. model
Source sentence
In order for the integrals defining M and V to converge, by Cauchy-Schwarz inequality it suffices to require that eF(t, θ, ·) ∈L2(Ω) and ∂uθ ∂θi ∈L2(Ω) for all t ∈[0, T], θ ∈Rp+1 and 1 ≤i ≤p + 1. It is convenient (but not necessary) to also demand that uθ ∈L2(Ω) for all θ ∈Rp+1.
Data
- Synthetic 3‑dimensional grid experiment synthetic unknown
Source sentence
The total space dimension including x, y is d = 3, and the space grid is [0, 4] along all dimensions. The default value of parameters are T = 1, γ = 0.5, r = 0.05, λ = 0.1, a = 0.3 0.0 0.0 0.3 , b = [0.2, 0.2] and ρ = [0.1, 0.1].
Costs
Not applicable
Finds 2
- For the non‑tradable stock example, the required integrals can be performed in closed form, yielding explicit ODEs for the variational parameters. theoretical
Source sentence
The necessary integrals can be performed in closed form and one finds the following results, M = α2 1 0 0 1 4 , V = α2 −1 2(λ2 + r) 1 4r . Plugging the above expressions into (2.2), we find the following time dependence for the variational parameters, d dt log α log β = −1 2(λ2 + r) r .
- The NG method reproduces the exact solution of the PDE in this case, as verified by substituting the trial function back into the original equation. theoretical
Source sentence
Remarkably, one finds that the NG reproduces the exact solution of the PDE in this case, which can be verified by plugging uθ(t) back into (2.1).
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 10 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (12)
- Normalizing flows for probabilistic modeling and inference (2021)
- Solving the quantum many-body problem with artificial neural networks (2017)
- DGM: A deep learning algorithm for solving partial differential equations (2018)
- Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations (2019) — link
- Variational Monte Carlo approach to partial differential equations with neural networks (2022)
- A deep learning algorithm for optimal investment strategies (2022)
- An Example of Indifference Prices Under Exponential Preferences (2004) — link
- A Valuation Algorithm for Indifference Prices in Incomplete Markets (2004) — link
and 4 more
- Optimal Hedging and Valuation of Nontraded Assets (2000) — link
- Valuation of Claims on Non-Traded Assets Using Utility Maximization (2002) — link
- Optimal Investment with Derivative Securities (2005) — link
- Quantum-inspired variational algorithms for partial differential equations: application to financial derivative pricing (2023) — link
0 of 12 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.