Paper: arXiv 2309.03966

Authors: Rong Du, Duy-Minh Dang

Abstract

This paper introduces FourNet, a novel single-layer feed-forward neural network (FFNN) method designed to approximate transition densities for which closed-form expressions of their Fourier transforms, i.e. characteristic functions, are available. A unique feature of FourNet lies in its use of a Gaussian activation function, enabling exact Fourier and inverse Fourier transformations and drawing analogies with the Gaussian mixture model. We mathematically establish FourNet’s capacity to approximate transition densities in the \(L_2\)-sense arbitrarily well with finite number of neurons. The parameters of FourNet are learned by minimizing a loss function derived from the known characteristic function and the Fourier transform of the FFNN, complemented by a strategic sampling approach to enhance training. We derive practical bounds for the \(L_2\) estimation error and the potential pointwise loss of nonnegativity in FourNet for \(d\)-dimensions (\(d\ge 1\)), highlighting its robustness and applicability in practical settings. FourNet’s accuracy and versatility are demonstrated through a wide range of dynamics common in quantitative finance, including Lévy processes and the Heston stochastic volatility models-including those augmented with the self-exciting Queue-Hawkes jump process.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper demonstrates high mathematical complexity through rigorous $L_2$-analysis, universal approximation theorems, and multi-dimensional error bounds, while maintaining strong empirical rigor with validated demonstrations across diverse financial models (Lévy, Heston, Queue-Hawkes) and practical error controls for real-world applications.

Research Flowchart

  flowchart TD
  A["Research Goal"] --> B["FourNet Architecture"]
  A --> C["Input Data"]
  subgraph B ["Neural Network Structure"]
      B1["Gaussian Activation"]
      B2["Fourier Transform"]
  end
  subgraph C ["Financial Models"]
      C1["Levy Processes"]
      C2["Heston Model"]
      C3["Queue-Hawkes Jumps"]
  end
  B --> D["Training Process"]
  C --> D
  D --> E["Loss Function"]
  D --> F["Strategic Sampling"]
  E --> G["Minimization"]
  F --> G
  G --> H["Key Outcomes"]
  subgraph H ["Findings"]
      H1["Arbitrary L2 Approximation"]
      H2["Pointwise Error Bounds"]
      H3["Preserves Nonnegativity"]
  end

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 proposes FourNet, a single‑layer feed‑forward neural network with a Gaussian activation that approximates transition densities using their known characteristic functions.

Source sentence
This paper introduces FourNet, a novel single-layer feed-forward neural network (FFNN) method designed to approximate transition densities for which closed-form expressions of their Fourier transforms, i.e. characteristic functions, are available.

Assumes 4

  • The parameter space Θ is bounded, implying bounded first derivatives of the learned Fourier transform. model
    Source sentence
    Given that Θ is bounded and thus bG(η; θ) possesses a bounded first derivative, the requirement for C′ < ∞ is that G(η) also has a bounded first derivative.
  • The random variable associated with the density g(·) is absolutely integrable. model
    Source sentence
    This leads us to the assumption that the random variable associated with the density g(·) is absolutely integrable.
  • The characteristic function G(·) and its learned counterpart bG(·;θ) are square‑integrable (L2) and integrable (L1) over the Fourier domain. statistical
    Source sentence
    Both |Re bG(·; θ)| and |Im bG(·; θ)|, θ ∈Θ, are in L1(R). We also recall that both |ReG(·)| and |ImG(·)| ∈L1(R)
  • The loss function includes an MAE regularization term that controls pointwise loss of non‑negativity. other
    Source sentence
    The MAE component enables direct control over the upper bound of the potential pointwise loss of non‑negativity in bg(·; θ).

Data

  • Synthetic transition‑density datasets synthetic unknown
    Source sentence
    All datasets feature one-dimensional inputs (d = 1), except for the two-dimensional Merton jump-diffusion process discussed in Section 6.3.

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • FourNet can approximate transition densities arbitrarily well in the L2‑sense with a finite number of neurons. theoretical
    Source sentence
    We mathematically establish FourNet’s capacity to approximate transition densities in the L2-sense arbitrarily well with finite number of neurons.
  • Training achieves a loss ≤10⁻⁶ and the MAE regularization term is also <10⁻⁶. performance
    Source sentence
    The parameters bθ is learned through training, satisfying LossP (bθ) ≤10−6. This implies that the MAE regularization term RP (θ)
  • The measured loss of non‑negativity of the estimated density is about 10⁻⁷, negligible for practical purposes. performance
    Source sentence
    the measure for loss of non-negativity | min(bg(·; bθ), 0)| is about 10−7, negligible for all practical purpos…
  • European and Bermudan option prices computed using FourNet‑estimated transition densities exhibit impressive accuracy. pricing-calibration
    Source sentence
    computed using estimated transition densities obtained through FourNet exhibit impressive accuracy.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 35 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive

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.