Riemann–Liouville Fractional Brownian Motion

• RL-fBm is a self-similar, zero-mean Gaussian process defined via a one-sided fractional integral of white noise, extending standard Brownian motion.
• It exhibits nonstationary increments, aging, and anomalous diffusion with variance scaling as t^(2H), differentiating it from classical FBM models.
• The process is pivotal in modeling anomalous transport and stochastic systems, offering unique insights through its covariance structure and ergodic properties.

Riemann-Liouville fractional Brownian motion (RL-fBm) is a zero-mean, self-similar, non-Markovian Gaussian process with long-range dependence and nonstationary increments, defined via a one-sided fractional integral of Gaussian white noise. RL-fBm generalizes standard Brownian motion and fractional Brownian motion (FBM) by extending the classical integration to a fractional order, indexed by the Hurst exponent $H \in (0,1)$. Its pathwise, covariance, and ergodic properties fundamentally distinguish it from the Mandelbrot–van Ness (MV) and Langevin equation forms of FBM, primarily due to its intrinsic non-stationarity and boundary condition at $t=0$ (Woszczek et al., 2024, Benichou et al., 2023, Wei et al., 2024, Wang et al., 27 Apr 2025, Brzezniak et al., 2010, Garra et al., 2018, McGonegal, 2014).

1. Mathematical Definition and Integral Representation

RL-fBm is constructed as a Volterra–type stochastic integral of white noise or Brownian motion with a causal Riemann–Liouville kernel. For $H \in (0,1)$,

$$B_H^{\mathrm{RL}}(t) = \frac{1}{\Gamma(H+\tfrac12)} \int_{0}^{t} (t-s)^{H-\tfrac12} \,\mathrm{d}W(s),$$

where $W$ is standard Brownian motion, and $\Gamma$ denotes the Gamma function. The increment kernel is strictly one-sided, integrating only over $[0, t]$, in contrast to MV-FBM, which symmetrizes over the infinite past and (optionally) future (Woszczek et al., 2024, Benichou et al., 2023, Wang et al., 27 Apr 2025, Brzezniak et al., 2010).

Formally, RL-fBm may also be viewed as the image of white noise under a fractional Riemann–Liouville integral $I_{0+}^{H+\tfrac12}$, i.e., $B_H^{\mathrm{RL}}(t) = I_{0+}^{H+\tfrac12} \xi(t)$. For $H = 1/2$, this reduces to standard Brownian motion (Benichou et al., 2023, McGonegal, 2014).

2. Covariance Structure and Self-similarity

RL-fBm is a centered Gaussian process $\mathbb{E}[B_H^{\mathrm{RL}}(t)] = 0$ with covariance: $$\mathbb{E}[B_H^{\mathrm{RL}}(t_1)B_H^{\mathrm{RL}}(t_2)] = \frac{1}{\Gamma(H+\tfrac12)^2} \int_{0}^{\min(t_1,t_2)} (t_1-s)^{H-\tfrac12} (t_2-s)^{H-\tfrac12} \, ds.$$ For the variance, set $t_1 = t_2 = t$ to obtain: $$\mathbb{E}[B_H^{\mathrm{RL}}(t)^2] = \frac{t^{2H}}{2H\,\Gamma(H+\tfrac12)^2} \propto t^{2H}.$$ The process is strictly self-similar, i.e., $$\{B_H^{\mathrm{RL}}(a t)\} \stackrel{\mathrm{law}}{=} \{a^H B_H^{\mathrm{RL}}(t)\}$$ for all $a > 0$, but unlike MV-FBM, the increments are nonstationary (Woszczek et al., 2024, Benichou et al., 2023, Wang et al., 27 Apr 2025, Brzezniak et al., 2010, McGonegal, 2014).

3. Increment Statistics, Nonstationarity, and Aging

Given $$\Delta B(t;\tau) = B_H^{\mathrm{RL}}(t+\tau) - B_H^{\mathrm{RL}}(t)$$, the mean squared increment (MSI) is given by

$$\mathbb{E}\left[\Delta B(t;\tau)^2\right] = \frac{1}{\Gamma(H+\tfrac12)^2} \left\{ \int_0^t \left[ (t+\tau-s)^{H-\tfrac12}-(t-s)^{H-\tfrac12} \right]^2 ds + \int_t^{t+\tau} (t+\tau-s)^{2H-1} ds \right\}.$$

Equivalently, this can be expressed as

$$\mathbb{E}\left[\Delta B(t;\tau)^2\right] = A_H \tau^{2H} \left[ I_H(t/\tau) + \frac{1}{2H} \right],$$

where

$$A_H = 1/\Gamma(H+\tfrac12)^2,\quad I_H(z) = \int_0^z \left[ (1+s)^{H-\tfrac12} - s^{H-\tfrac12} \right]^2 ds.$$

For small $t \ll \tau$, the process behaves similarly to standard FBM ($\tau^{2H}$ scaling), but for large $t \gg \tau$, the variance approaches a stationary form with a prefactor dependent on $H$, different from that of MV-FBM (Woszczek et al., 2024, Wang et al., 27 Apr 2025).

The time-averaged mean squared displacement (TAMSD) for a single trajectory of length $T$ is

$$\delta_H^*(\tau) = \frac{1}{T-\tau} \int_0^{T-\tau} [B_H^{\mathrm{RL}}(t+\tau) - B_H^{\mathrm{RL}}(t)]^2 dt,$$

whose ensemble mean, in the limit $T \gg \tau$, scales as $\mathbb{E}[\delta_H^*(\tau)] \sim C(H) \tau^{2H}$, but with a different prefactor from MV-FBM: $$C(H) = \frac{2H \Gamma(H+\tfrac12)^2}{\Gamma(1+2H)\,\sin(\pi H)}.$$ This scaling underlies its classification as exhibiting anomalous diffusion with exponent $2H$ (Woszczek et al., 2024, Wei et al., 2024).

RL-fBm exhibits explicit "age" or initial-time dependence in finite-$t$ observables: the increment variance and autocovariance functions depend on both lag $\tau$ and starting time $t$, a phenomenon sometimes termed "aging" (Wang et al., 27 Apr 2025, Wei et al., 2024).

4. Ergodicity, Stationarity, and Memory

Although RL-fBm is Gaussian and self-similar, its increments are not generally stationary. For $H \in (0,1)$, and particularly $1/2 < \alpha < 3/2$ with $H = \alpha - 1/2$, the process has asymptotically stationary increments as $t \gg \tau$, but the ensemble mean TAMSD and MSD do not coincide: $$\lim_{T\to\infty,\,\tau/T\to0} \overline{\delta^2(\tau)} = S(\tau) \neq \langle B_H^{\mathrm{RL}}(\tau)^2 \rangle.$$ This "spurious nonergodicity" is rectified under strong aging ($t_a \gg t, T$), where the increment statistics become stationary and ergodicity is restored (Wei et al., 2024).

Higher-order increments (finite differences) can exhibit restored stationarity in the increments when the order is high enough, specifically when $(2n+1)/2 < \alpha < (2n+3)/2$ (Wei et al., 2024).

RL-fBm is not mixing, and classical ergodic theorems do not apply due to nonstationary increments. Its long-range memory persists: for fixed lag $\tau$, $$\mathrm{Cov}(B_H^{\mathrm{RL}}(t), B_H^{\mathrm{RL}}(t+\tau)) \sim t^{2H}$$ as $t \to \infty$ (Woszczek et al., 2024, Wei et al., 2024).

5. Path Properties and Stochastic Calculus

RL-fBm is almost surely Hölder continuous with any exponent $\gamma < H$. The process is nowhere differentiable, with unbounded $p$-variation unless $p > 1/H$; quadratic variation vanishes for $H > 1/2$ and diverges for $H < 1/2$. RL-fBm is not a (semi-)martingale unless $H = 1/2$ (McGonegal, 2014, Brzezniak et al., 2010).

Stochastic integration with respect to RL-fBm differs substantially from the standard Itô theory. For $0 < H < 1/2$, the integration theory for operator-valued integrands is equivalent—up to normalization constants—to integration with respect to classical FBM, especially in the context of stochastic evolution equations. The RKHS structure established by Riemann–Liouville fractional integrals is essential for analyzing regularity and solvability of SPDEs (Brzezniak et al., 2010).

6. Relation to Fractional Langevin Dynamics and Generalized Diffusion Equations

RL-fBm arises naturally as the solution to certain fractional Langevin equations with Caputo derivative of order $\alpha$, i.e.,

$${}^C_0 D_t^\alpha X(t) = \sqrt{2 K_\alpha}\,\xi(t),$$

with $H = \alpha - 1/2$, and initial conditions set to ensure RL-type causality (Wei et al., 2024). The process also emerges as the fundamental solution to generalized diffusion equations with time-dependent diffusivity $K(t)$ obeying a nonlinear Riemann–Liouville evolution equation: $D_{0+}^{1-2H} K(t) = k K^2(t),$ and $K(t) = 2H C t^{2H-1}$, yielding variance scaling $\sim t^{2H}$ (Garra et al., 2018).

This establishes RL-fBm as a canonical model for anomalous diffusion in systems with long-time memory and non-equilibrium effects, where standard fluctuation-dissipation balance may not hold (Wei et al., 2024, Garra et al., 2018).

7. Comparison with Other FBM Constructions

The key distinctions between RL-fBm and alternative FBM definitions are summarized below:

Feature	RL-fBm	MV-fBm / Standard FBM	LE-FBM
Kernel	One-sided, $(t-s)^{H-1/2}$ over $[0,t]$	Two-sided, includes past and future	Langevin with correlated noise
Increments	Nonstationary	Stationary	Stationary
Covariance	Integral involving hypergeometric	Closed form (power law in lags)	As for MV-fBm
Ergodicity	No (spurious nonergodicity)	Yes	Yes
Ageing	Explicit t-dependence	None (time-homogenous)	None
Path regularity	Hölder $< H$	Hölder $< H$	Hölder $< H$

Only RL-fBm displays explicit dependence on the initial time in its two-point and higher-order statistics, resulting in observable aging phenomena and unique nonstationary nonergodicity signatures (Wang et al., 27 Apr 2025, Wei et al., 2024).

8. Applications and Relevance

RL-fBm serves as a universal model for systems exhibiting anomalous transport and aging—critical in biological, soft-matter, and out-of-equilibrium statistical physics contexts. Its distinct nonstationarity provides a means to distinguish between models and experimental systems with otherwise similar anomalous diffusion scaling but different time- and sample-averaged behaviors (Woszczek et al., 2024, Wei et al., 2024, Wang et al., 27 Apr 2025).

Its kernel representation underlies both stochastic calculus (Malliavin, pathwise) and fractional Fokker–Planck formulations. Applications extend to stochastic partial differential equations driven by RL-fractional temporal noise, with established regularity and solution theory in both Hilbert and Banach spaces (Brzezniak et al., 2010). The RL-fBm formulation accommodates parameter randomization (random Hurst indices) essential for modeling heterogeneous environments and heterogeneous diffusive regimes (Woszczek et al., 2024).

---

References:

(Woszczek et al., 2024, Benichou et al., 2023, Wei et al., 2024, Wang et al., 27 Apr 2025, Brzezniak et al., 2010, Garra et al., 2018, McGonegal, 2014)