Fractional Brownian motion

Fractional Brownian motion, commonly abbreviated fBm, is a centered Gaussian process whose covariance is determined by a parameter (H\in(0,1)), called the Hurst parameter. It generalizes Brownian motion while preserving Gaussianity, self-similarity, and stationary increments. Except when (H=\tfrac12), it has dependent increments and is neither a Markov process nor a semimartingale.

A standard fractional Brownian motion ({B_H(t):t\geq 0}) satisfies

[ \mathbb E[B_H(t)]=0 ]

and

[ \mathbb E[B_H(t)B_H(s)]

\frac12\left( t^{2H}+s^{2H}-|t-s|^{2H} \right). ]

Consequently,

[ \mathbb E!\left[(B_H(t)-B_H(s))^2\right]

|t-s|^{2H}. ]

The exponent (2H) controls the scaling of increment variance and the local regularity of sample paths. When (H=\tfrac12), the covariance reduces to (\min(s,t)), so the process is ordinary standard Brownian motion.

Historical formulation

Power-law Gaussian processes with covariance structures related to fractional Brownian motion appeared in the work of Andrey Kolmogorov during the 1940s. The modern formulation was established in 1968 by Benoît Mandelbrot, John W. Van Ness, and You Watanabe, who connected the covariance model to a two-sided moving average of Wiener noise and adopted the term “fractional Brownian motion.” Their construction clarified how the parameter (H) produces either persistent or antipersistent dependence while retaining a Gaussian scaling law.

The two-sided integral representation developed by Mandelbrot and Van Ness also linked the process to fractional calculus. The word “fractional” refers to the fractional-order integration implicit in this representation rather than to noninteger time or non-Gaussian probability distributions.

Defining properties

Fractional Brownian motion is (H)-self-similar. For every (a>0),

[ {B_H(at):t\geq0} \overset{d}{=} {a^H B_H(t):t\geq0}, ]

where equality holds in the sense of all finite-dimensional distributions. This scaling relation implies that magnifying the time coordinate by (a) magnifies typical displacement by (a^H).

The increments are stationary because the distribution of

[ B_H(t+h)-B_H(t) ]

depends on (h) but not on (t). They are generally not independent, even when they are taken over disjoint intervals. Independence occurs at (H=\tfrac12), where Gaussianity and zero covariance together recover the independent increments of Brownian motion.

For unit increments

[ X_k=B_H(k+1)-B_H(k), ]

the resulting stationary Gaussian sequence is called fractional Gaussian noise. Its autocovariance is

[ \gamma(k)

\frac12\left( |k+1|^{2H} -2|k|^{2H} +|k-1|^{2H} \right). ]

At large lags,

[ \gamma(k)\sim H(2H-1)|k|^{2H-2}. ]

When (H>\tfrac12), distant increments have positive covariance, and the covariance series is not absolutely summable. This regime exhibits long-range dependence, with changes over one interval statistically aligned with changes over later intervals. When (H<\tfrac12), nonoverlapping increments are negatively correlated, producing antipersistent dependence whose correlations decay sufficiently rapidly to be summable. At (H=\tfrac12), all nonzero-lag increment covariances vanish.

Integral representations

For a two-sided standard Brownian motion (W), fractional Brownian motion can be represented for (t\geq0) as

[ B_H(t)

C_H \left[ \int_{-\infty}^{0} \left( (t-s)^{H-\frac12}

(-s)^{H-\frac12} \right),dW(s) + \int_{0}^{t} (t-s)^{H-\frac12},dW(s) \right], ]

where the positive constant (C_H) is selected so that (\mathbb E[B_H(1)^2]=1). The subtraction in the first integral compensates for the contribution from the remote past and ensures convergence. At (H=\tfrac12), the kernel reduces to the ordinary Wiener construction after cancellation of the past term.

A harmonizable representation expresses the same process through frequency-domain Gaussian noise:

[ B_H(t)

\widetilde C_H \int_{\mathbb R} \frac{e^{it\xi}-1}{|\xi|^{H+\frac12}} ,\widehat W(d\xi), ]

where (\widehat W) is a suitably symmetric complex Gaussian random measure. The factor (e^{it\xi}-1) enforces (B_H(0)=0), while the power of (|\xi|) generates the required scaling exponent. This representation connects fractional Brownian motion with spectral analysis, generalized random fields, and homogeneous Gaussian structures.

Path regularity

Fractional Brownian motion has a continuous modification. With probability one, its sample paths are locally Hölder continuous of every order (\alpha<H). They are almost surely not locally Hölder continuous of order (H), so the Hurst parameter gives the critical regularity exponent.

For every (H\in(0,1)), the paths are almost surely nowhere differentiable. Increasing (H) nevertheless produces greater local regularity in the Hölder sense. The graph over a bounded time interval has almost-sure Hausdorff dimension

[ 2-H, ]

which equals (\tfrac32) in the Brownian case and decreases as the paths become more regular.

The quadratic variation differs sharply across parameter regimes. Along refining regular partitions, it diverges when (H<\tfrac12), converges to elapsed time when (H=\tfrac12), and vanishes when (H>\tfrac12). This behavior is one manifestation of the failure of the ordinary semimartingale framework outside the Brownian case.

Stochastic integration

Because fractional Brownian motion is not a semimartingale for (H\neq\tfrac12), the standard Itô calculus does not apply to it without modification. Several inequivalent integration theories are used because each encodes a different relation between the integrand and the Gaussian driver.

For sufficiently regular integrands and (H>\tfrac12), integration can be defined pathwise through Young integration. The construction depends on complementary Hölder regularity and does not require a martingale structure. For rougher parameter ranges, rough path theory augments the process with iterated integrals that retain information lost at the level of the path alone.

A separate probabilistic construction uses the divergence operator from Malliavin calculus. The resulting Skorohod-type integral need not agree with a pathwise integral, particularly when the integrand depends on the future or is otherwise correlated with the process. Formulae resembling the classical chain rule therefore depend on the chosen integration framework.

Statistical and modeling role

Fractional Brownian motion is the unique centered Gaussian process, up to a multiplicative scale factor, that has stationary increments and is self-similar with exponent (H). This characterization explains its role as a canonical Gaussian model for scale-invariant fluctuations with temporal dependence.

Its increment process is used in mathematical models where correlations follow a power law rather than an exponential decay. Such models occur in analyses of network traffic, hydrological records, material surfaces, and anomalous diffusion. The Gaussian assumption fixes all finite-dimensional distributions through the covariance, so the model does not separately represent heavy-tailed marginal behavior or discontinuous jumps.

Parameter estimation commonly relies on how increment variance changes with temporal scale, on the low-frequency spectral behavior of fractional Gaussian noise, or on the Gaussian likelihood associated with its covariance matrix. Finite samples can exhibit trends and nonstationary components that produce scaling patterns resembling those generated by fractional dependence, making the separation of deterministic structure from stochastic scaling part of the statistical formulation.

Related processes

Fractional Brownian motion differs from Lévy processes because its increments are generally dependent, although both classes contain Brownian motion as a special case. It also differs from fractional Lévy motion, which replaces the Gaussian driving noise with a Lévy source and may consequently have heavy tails or jumps.

The Ornstein–Uhlenbeck process has short-range dependence and a characteristic relaxation time, whereas fractional Brownian motion is scale-free and nonstationary. A stationary mean-reverting process can instead be formed by driving a linear relaxation equation with fractional Brownian motion; the resulting model is called a fractional Ornstein–Uhlenbeck process.

Multifractional Brownian motion replaces the constant (H) with a function of time. This modification allows local regularity to vary, but it generally removes exact self-similarity and stationary increments.

See also

  • Brownian motion, the (H=\tfrac12) member of the fractional Brownian family and the Gaussian process underlying classical Itô calculus.
  • Fractional Gaussian noise, the stationary increment sequence obtained by sampling fractional Brownian motion at equally spaced times.
  • Hurst exponent, the scaling parameter that determines self-similarity, path regularity, and increment dependence.
  • Long-range dependence, the nonsummable correlation structure exhibited by fractional Gaussian noise when (H>\tfrac12).
  • Fractional calculus, the theory of noninteger-order operators associated with the moving-average kernels of fractional processes.
  • Rough path theory, a framework for differential equations driven by irregular signals, including fractional Brownian motion in suitable parameter ranges.
  • Multifractional Brownian motion, a Gaussian extension in which the local Hurst parameter varies with time.