Fractional Brownian Motion Overview

• Fractional Brownian motions are self-similar Gaussian processes defined by the Hurst exponent, capturing long-range dependence and non-Markovian behavior.
• They provide a unifying framework for diverse constructions, such as the Lévy and Mandelbrot–van Ness approaches, distinguished by their covariance and increment properties.
• Their analytical flexibility supports simulation, path-integral representations, and extensions to multivariate or higher-order processes in modeling anomalous diffusion.

Fractional Brownian Motions (fBms) are a fundamental family of self-similar Gaussian processes that generalize classical Brownian motion by incorporating long-range dependence and non-Markovian structure via a continuous parameter: the Hurst exponent. Originally formalized independently by Lévy and Mandelbrot–van Ness (MvN), fBms are distinguished by their rich covariance structure, pivotal roles in modeling anomalous diffusion, and wide applicability in probability theory, statistical physics, and fields with memory effects.

1. Alternative Constructions and Covariance Structures

Fractional Brownian motions are defined as centered Gaussian processes $\{X(t)\}$ uniquely determined by their covariance function, parameterized by $H \in (0,1)$ (the Hurst exponent):

• Lévy fBm (left-sided Riemann–Liouville construction): Defined for $t\in[0,T]$ as

$$x_L(t) = \frac{1}{\Gamma(H+\frac{1}{2})} \int_0^t (t-s)^{H-\frac{1}{2}}\ \xi(s)\ ds,$$

where $\xi(s)$ is standard Gaussian white noise. The covariance has the closed form

$$\langle x_L(t_1) x_L(t_2)\rangle = (H+\tfrac{1}{2}) T^{2(H+\frac{1}{2})} (t_1 t_2)^{\frac{1}{2}-H} {}_2F_1(1,2-H;H+1;t_2/t_1)$$

where ${}_2F_1$ is the Gauss hypergeometric function.

• Mandelbrot–van Ness (MvN) fBm (Weyl integral): For two-sided time $t\in(-\infty,\infty)$,

$$x_{MvN}(t) = C_H\left[ \int_{-\infty}^t (t-s)^{H-\frac{1}{2}} \xi(s) ds - \int_{-\infty}^0 (-s)^{H-\frac{1}{2}} \xi(s) ds \right].$$

The covariance for the two-sided version is

$$C_{MvN}(t_1,t_2) = \tfrac{1}{2}\left(|t_1|^{2H} + |t_2|^{2H} - |t_1 - t_2|^{2H}\right),$$

which for $t_1,t_2\ge0$ also defines the one-sided MvN fBm.

Construction	Time Domain	Covariance Structure	Increment Stationarity
Lévy (RL integral)	$[0,T]$	Hypergeometric function, non-stationary increments	No (depends on $t_1$, $t_2$)
MvN (Weyl integral)	$(-\infty,\infty)$	Power-law, stationary increments	Yes (depends only on $t_1-t_2$)

The MvN construction yields stationary increments, while the Lévy form does not. Nevertheless, the scaling of variance is $Var[X(t)] \sim t^{2H}$ in all cases (Benichou et al., 2023).

2. Path-Integral Representations and Unifying Framework

Historically, the physical and mathematical literature featured distinct path-integral ("action") representations for each construction, leading to debate over their equivalence:

• Lévy case (subdiffusive, $0

$$S_L[x] = \frac{1}{2}\int_0^T [ D_{0+}^{H+\frac{1}{2}} x(t) ]^2 dt,$$

where $D_{0+}^\alpha$ is a left-sided Riemann–Liouville fractional derivative.

• MvN case (e.g., two-sided, subdiffusive):

$$S_{MvN}[x] = \frac{1}{2}\int_{-\infty}^{\infty} \int_{-\infty}^{\infty} \dot{x}(t_1) \dot{x}(t_2) \frac{c_H}{|t_1-t_2|^{2H}} dt_1 dt_2$$

Despite differing analytical forms, all three definitions admit a unified representation in terms of Riemann–Liouville fractional integrals (Benichou et al., 2023):

• For all three cases, the quadratic action can be written as

$$S[x] = \frac{1}{2} \int_{T_-}^{T_+} \left[ I_{a}^{\nu} \dot{x}(t) \right]^2 dt$$

where $I_{a}^{\nu}$ is a (left or right) Riemann–Liouville fractional integral of order

$$\nu = \begin{cases} H + \tfrac{1}{2}, & H < 1/2, \ \frac{3}{2} - H, & H > 1/2, \end{cases}$$

and the domain/limits $(T_-,T_+,a)$ depend on the specific construction (Lévy: $[0,T],a=0$; one-sided MvN: $[0,\infty),a=\infty$; two-sided MvN: $(-\infty,\infty),a=-\infty$).

This equivalence demonstrates that all three are manifestations of the same Gaussian process, differing only by their domains and associated boundary conditions.

3. Self-Similarity, Stationarity, and Increment Structure

The central features of fBm, across all definitions:

• Self-similarity: For all $c > 0$,

$\{ X(ct) \} \stackrel{d}{=} \{ c^H X(t) \},$

with $H$ the Hurst exponent, irrespective of the construction.

• Increment Stationarity:
  • MvN fBms have stationary increments: $\operatorname{Cov}(X(t+h)-X(t))$ depends only on $h$.
  • Lévy fBms do not generally have stationary increments, but all have equivalent one-dimensional marginals and scaling.
• Anomalous Diffusion: The scaling

$\mathrm{Var}\,[X(t)] \sim t^{2H}$

interpolates between subdiffusive $(0 < H < 1/2)$, classical $(H = 1/2)$, and superdiffusive $(1/2 < H < 1)$ regimes.

• Covariances unify via the master action's structure, with all increment and path regularity properties following from the shared path measure (Benichou et al., 2023).

4. Extensions, Simulation, and Analytical Implications

The unifying representation has several consequences:

• Simulation: One can discretize the Riemann–Liouville kernel over the desired time interval instead of separately implementing each construction.
• Analysis: Analytical questions (e.g., first-passage times, large deviations, path conditioning) can be addressed generally using the master action and then specialized as needed for domain-specific constraints.
• Higher-Order and Multivariate Extensions: The structural unity allows construction of fBms with higher-order time derivatives or multi-dimensional generalizations (operator fractional Brownian motions and anisotropic Hurst indices) by varying the order of fractional integration and integration domain (Benichou et al., 2023).

5. Implications and Unified Gaussian Family

Despite superficial differences, Lévy and Mandelbrot–van Ness fractional Brownian motions are not distinct processes but specializations of a single Gaussian family with a common analytical foundation:

• All are defined by the same class of quadratic Gaussian measures in path space, parametrized by the choice of Riemann–Liouville kernel and integral bounds.
• The increment-stationarity and covariance distinctions are entirely attributable to the boundary conditions of the underlying domain, not to distinct stochastic structures.
• This insight simplifies theoretical, numerical, and physical modeling, reducing ambiguity in choosing among fBm variants for modeling anomalous diffusion or long-memory processes (Benichou et al., 2023).

• Key unification and path-integral framework: O. Bénichou & G. Oshanin, "A unifying representation of path integrals for fractional Brownian motions" (Benichou et al., 2023).

---

This synthesis establishes that all common constructions of fractional Brownian motion—Lévy, one-sided Mandelbrot–van Ness, and two-sided Mandelbrot–van Ness—form a single self-similar Gaussian process family, with differences arising exclusively from time domain restrictions and not from intrinsic stochastic distinctions.