Spectral Representation and Simulation of Fractional Brownian Motion

Spectral Representation and Simulation of Fractional Brownian Motion

Introduction and Problem Formulation

The paper "Spectral Representation and Simulation of Fractional Brownian Motion" (2412.12207) addresses the problem of constructing efficient, analytically tractable, and numerically robust spectral representations of fractional Brownian motion (fBm), a centered Gaussian process with covariance function

$$R_H(t, \tau) = \frac{1}{2}\left(t^{2H} + \tau^{2H} - |t-\tau|^{2H}\right)$$

for Hurst index $H \in (0, 1)$ and $t, \tau \geq 0$. The fBm and its associated stochastic calculus extend Brownian-driven models, capturing long-memory and self-similarity effects relevant in stochastic differential equations (SDEs), turbulence, hydrology, finance, and telecommunications.

Traditional simulation of fBm relies on integral representations involving Wiener processes and nontrivial kernels expressed through hypergeometric functions, wavelets, or trigonometric series. However, these approaches often result in cumbersome numerical procedures, nontrivial convergence analysis, or lose accuracy/resilience to high-dimensional truncations.

The author proposes a new representation based on the spectral form of mathematical description: spectral expansion over Legendre polynomials, leading to an orthogonal basis in $L^2([0, T])$. This choice enables explicit, easily computable formulas for the expansion coefficients, leverages favorable numerical properties of Legendre systems, and makes it possible to analyze and bound approximation errors directly.

Spectral Methodology

The central technical advance is the application of the spectral method—expanding fBm and associated linear operators into a Legendre polynomial basis, leading to representations in terms of infinite (or truncated finite) matrices of spectral coefficients. Specifically, the approach includes:

• Spectral Characteristics of Functions and Operators: For any function $f \in L^2([0, T])$, the spectral characteristic is the column vector of expansion coefficients in the Legendre basis. For kernels $k(s, t)$ representing Hilbert–Schmidt operators, the spectral characteristic is the infinite matrix $\{K_{ij}\}$ where

$$K_{ij} = \langle \hat{P}_i \otimes \hat{P}_j, k \rangle.$$

• Decomposition of the fBm Kernel: Starting from an integral operator representation for fBm,

$B_H(t) = \int_0^t k_H(t, \tau)\, dB(\tau),$

where $B(\cdot)$ is classical Brownian motion and $k_H$ is an operator-valued kernel, the author adopts the Decreusefond-Üstünel kernel representation, which is then explicitly decomposed in the Legendre basis.

• Spectral Factorizations: The fBm-shaping operator $K_H$ is factorized as products of spectral characteristics of (generalized) multiplication by $t^\alpha$ and fractional integration operators $J^\beta_{0+}$, with the cases $H < 1/2$ and $H > 1/2$ addressed separately. The construction ensures spectral operators retain the operator-theoretic composition structure.
• Explicit Recurrences and Analytical Formulas: For the spectral characteristics, the author derives compact recurrence relations for the Legendre expansion of power functions and for the multiplication and fractional integration operators in the Legendre basis. This yields, for example,

$$F_i^\alpha = T^\alpha \sqrt{T} \frac{\sqrt{2i+1} \alpha (\alpha-1)\cdots(\alpha - i + 1)}{(\alpha+1)(\alpha+2)\cdots(\alpha+i+1)}$$

for the expansion coefficients of $t^\alpha$.

• Algorithmic Stability: Exploiting the sparse and structured nature of Legendre basis expansions, the author develops numerically stable algorithms capable of achieving accurate results for matrices of high dimension ($L$ up to 256), mitigating the numerical instability typically found in naive power expansions or with trigonometric/wavelet bases.

Numerical Analysis and Simulation Results

A rigorous analysis of approximation errors, convergence rates, and computational aspects is presented. Key findings include:

• Error Quantification: The total $L^2$ error between the true and approximated covariance functions can be split into truncation error (from spectral truncation) and computational error (from finite-precision calculation of matrix elements). Explicit formulas for both are provided.
• Empirical Accuracy and Convergence: Across Hurst parameters $H \in (0.1, 0.9)$ and increasing spectral truncation order $L$, the method exhibits convergence rates $\gamma$ between 1.3 (worst case) and over 3.2. For moderate $L$ (e.g., $L=64$), the relative error in covariance can be reduced below $0.2\%$ for $H$ in $(0.4, 0.7)$.
• Comparison with Other Bases: For $H = 1/2$ (classical Brownian motion), where exact analytic forms for covariance are known in multiple bases, Legendre polynomials yield lower approximation errors than trigonometric (Fourier or even cosine) or Walsh/Haar basis sets for the same truncation order, reflecting their efficiency for this class of Volterra-type kernels.
• Sample Path Generation: The method supports the generation of strong approximations to fBm sample paths via sampled spectral coefficients with independent Gaussian weights, directly realizing the prescribed covariance and temporal regularity properties.

Theoretical Implications

The research provides several important implications:

• Operator-theoretic Insights: The explicit connection between the structure of the fBm kernel, fractional integration, and spectral characteristics clarifies the functional-analytic nature of fBm simulation. It also establishes a template for treating other non-Markovian Gaussian processes with nontrivial covariance and similar Volterra structures.
• Modularity: The factorization of the fBm simulation operator in terms of spectral characteristics allows reuse of its components (e.g., multiplication operators and fractional integration operators) in other numerical methods for stochastic (especially fractional or non-Markovian) differential equations.
• Direct Computational Error Control: The framework allows direct quantification and control of the approximation error for covariance functions and, by extension, distributions of the simulated processes.
• Generalizability: This spectral approach is adaptable to related processes such as Liouville fractional Brownian motion and can be expected to integrate well with higher-order simulation algorithms for stochastic calculus.

Future Directions

The analytic and algorithmic results open several avenues:

• Extension to multi-dimensional fBm and operator-valued formulations in Hilbert-space settings, supporting high-dimensional or SPDE models.
• Application to SDEs (or SPDEs) driven by fBm, particularly for constructing high-order Taylor-type schemes based on spectral Itô/Stratonovich integrals.
• Investigation of optimal basis selection for more exotic covariance structures and more general classes of self-similar or long-memory processes.
• GPU/parallel implementation leveraging the inherent structure of spectral matrix operations for large-scale simulation scenarios.

Conclusion

This work provides a comprehensive and technically rigorous spectral representation of fractional Brownian motion in terms of Legendre polynomial expansions, with explicit analytical and algorithmic formulations for the associated operators and basis functions. The approach achieves robust, controllable accuracy for simulation and analysis of fBm, outperforms standard bases in terms of convergence and numerical stability, and yields reusable methodology for the numerical analysis of SDEs with memory. The modular spectral framework has significant implications for theory and simulation of non-Markovian processes, with ready application to advanced stochastic modeling and numerical methods.