--- title: '2D fBm: Theory, Geometry & Scaling' url: https://www.emergentmind.com/topics/two-dimensional-fractional-brownian-motion-2d-fbm type: topic --- # 2D fBm: Theory, Geometry & Scaling Two-dimensional fractional Brownian motion (2D fBm) is a vector-valued Gaussian process that extends one-dimensional fractional Brownian motion to planar trajectories while retaining self-similarity, long-range temporal dependence, and non-Markovianity. In its classical isotropic form, it is built from two Cartesian fBm components with a common Hurst exponent \(H\in(0,1)\), either independent or, in more general formulations, cross-correlated through a spatial covariance structure or correlated driving noises. Across the literature, 2D fBm appears in several complementary roles: as a Gaussian field with stationary increments, as a path measure with nonlocal action, as a rough path with Lévy area, and as a model for anomalous exploration, first passage, and anisotropic diffusion [1306.1667][2407.11655][2509.07537]. ## 1. Definitions and covariance structures The standard multi-dimensional construction defines fractional Brownian motion as a superposition of one-dimensional FBMs along each Cartesian coordinate, \[ \mathbf{x}_\alpha(t)=\sum_{i=1}^d \int_0^t \xi_\alpha^{(i)}(t')\,\hat{x}_i, \] where the \(\xi_\alpha^{(i)}\) are independent copies of fractional Gaussian noise, \(0<\alpha<2\), and \(H=\alpha/2\). In two dimensions, \[ \mathbf{x}_\alpha(t)=\big(x_\alpha^{(1)}(t),x_\alpha^{(2)}(t)\big), \] with independent components, zero mean, and mean-square displacement \[ \langle |\mathbf{x}_\alpha(t)|^2\rangle = 4K_\alpha t^\alpha. \] Each component has covariance \[ \langle x_\alpha^{(i)}(t_1)x_\alpha^{(i)}(t_2)\rangle =K_\alpha\big(|t_1|^\alpha+|t_2|^\alpha-|t_1-t_2|^\alpha\big), \] while \(\langle x_\alpha^{(1)}x_\alpha^{(2)}\rangle=0\). In this isotropic setting, the absolute magnitude \(|\mathbf{x}_\alpha(t)|\) is *not* a one-dimensional FBM, a point that becomes important in radial first-passage problems [1306.1667]. A closely related \(d\)-dimensional formulation uses a centered Gaussian process \(\mathbf{x}(t)=(x_1(t),\dots,x_d(t))\) with stationary increments and covariance \[ \operatorname{Cov}[x_i(t),x_j(t')] =\delta_{ij}\,D\big(t^{2H}+t'^{2H}-|t-t'|^{2H}\big). \] For \(d=2\), this yields the same planar model with independent, identically distributed coordinates and typical displacement \(\|\mathbf{x}(t)\|\sim t^H\) [2407.11655]. More general 2D fBm models allow dependent components. A recent construction introduces correlated Gaussian noises \(\tilde W_1,\tilde W_2\) with \[ \langle d\tilde W_1(t)\,d\tilde W_2(t)\rangle=\rho\,dt,\qquad |\rho|\le 1, \] and defines a two-component process \(\mathbf{X}(t)=(X_1(t),X_2(t))^\top\) by Mandelbrot–van Ness-type stochastic integrals. In the causal version, each marginal is exactly a standard Mandelbrot–van Ness fBm with its own Hurst index \(H_1,H_2\in(0,1)\), while dependence enters only through the noise correlation. The corresponding Hurst operator is diagonal, \[ H=\operatorname{diag}(H_1,H_2), \] so anisotropic scaling is componentwise: \[ \{\mathbf{X}(ct)\}_{t\in\mathbb{R}} \stackrel{d}{=} \{c^H\mathbf{X}(t)\}_{t\in\mathbb{R}},\qquad c^H=\operatorname{diag}(c^{H_1},c^{H_2}). \] The general covariance takes the form \[ \gamma_{jk}(t,s)=\frac{\sigma_j\sigma_k}{2}\Big[ w_{jk}(t)|t|^{H_j+H_k} +w_{jk}(-s)|s|^{H_j+H_k} -w_{jk}(t-s)|t-s|^{H_j+H_k} \Big], \] with \(w_{jk}\) determined by a cross-correlation parameter \(\rho_{jk}\) and, in the causal case, an asymmetry parameter \(\eta_{jk}\). The well-balanced version sets all \(\eta_{jk}=0\), producing a time-reversible process. If \(H_1=H_2\), the causal and well-balanced constructions have the same law [2509.07537]. ## 2. Self-similarity, scaling, and exploration regimes For the isotropic model, 2D fBm is self-affine: \[ \mathbf{x}_\alpha(\gamma t)\overset{\text{law}}{=}\gamma^{\alpha/2}\mathbf{x}_\alpha(t), \qquad H=\alpha/2. \] Its walk dimension satisfies \[ d_w=\frac{2}{\alpha}=\frac{1}{H}, \] and the typical radial displacement scales as \(t^H\). These exponents control the distinction between compact and non-compact exploration. In two dimensions, compact or recurrent exploration occurs for \(\alpha\le 1\), equivalently \(H\le 1/2\), while non-compact or transient exploration occurs for \(\alpha>1\), equivalently \(H>1/2\). Standard Brownian motion corresponds to the marginal case \(H=1/2\) [1306.1667][2407.11655]. A particularly useful summary parameter is \[ \mu=dH, \] which in two dimensions reduces to \[ \mu=2H. \] This single exponent organizes recurrence and territory coverage. For \(\mu<1\), 2D fBm is recurrent; for \(\mu=1\), marginally recurrent; for \(\mu>1\), transient. Thus the line \(H=1/2\) separates recurrent and transient planar motion. In the recurrent and marginal regimes, the number of distinct visited lattice cells scales as \[ \mathcal{N}(t)\sim t^{2H}, \] and the time needed to discover \(n\) distinct cells scales as \[ T_n\sim n^{1/(2H)}. \] In the transient regime, the same typical extent governs the explored territory, but many sites remain unvisited [2407.11655]. A standard fBm relation quoted in the visitation analysis gives the path fractal dimension as \[ d_{\text{path}}=\min(d,1/H). \] In two dimensions this implies \(d_{\text{path}}=2\) for \(H<1/2\), whereas for \(H\ge 1/2\) one has \(d_{\text{path}}=1/H\le 2\). This is consistent with the dense visited territory in recurrent 2D fBm and the sparser exploration in the superdiffusive transient regime [2407.11655]. ## 3. Gaussian measures and path-integral representations Because fBm is Gaussian, its law can be written as a quadratic functional. For a vector Gaussian process \(\mathbf{x}(t)\), \[ S[\mathbf{x}] =\frac12\int dt_1\int dt_2\,\mathbf{x}^\top(t_1)\mathbf{K}(t_1,t_2)\mathbf{x}(t_2), \qquad P[\mathbf{x}]\propto e^{-S[\mathbf{x}]}, \] with \(\mathbf{K}\) the inverse covariance kernel. For one-dimensional fBm, exact path-integral representations are available in terms of fractional Gaussian noise \(y(t)=\dot x(t)\) for \(00. \] In both regimes, the kernel is nonlocal in time and encodes the long-memory structure of fBm [2209.11722]. A unifying representation rewrites the action in terms of Riemann–Liouville fractional operators. In the subdiffusive case the action is a quadratic form of a fractional derivative of order \(1/2-H\) applied to \(dx/dt\); in the superdiffusive case it is a quadratic form of a fractional integral of order \(3/2-H\) applied to \(d^2x/dt^2\). The order of the operator depends only on whether \(H<1/2\) or \(H>1/2\); changing between Lévy, one-sided Mandelbrot–van Ness, and two-sided Mandelbrot–van Ness definitions modifies only the integration limits. A standard extension to 2D sums the componentwise actions for independent coordinates or inserts \(\Sigma^{-1}\) for correlated coordinates [2310.01863]. These actions also admit forced versions. For subdiffusive dynamics driven by \[ \dot{\mathbf{x}}(t)=\mathbf{f}[\mathbf{x}(t)]+\mathbf{y}(t), \] the natural vector action is \[ S[\mathbf{x}] = \frac12\int dt_1\int dt_2\, \big(\dot{\mathbf{x}}(t_1)-\mathbf{f}[\mathbf{x}(t_1)]\big)^\top \big[\mathcal{C}(t_1-t_2)\Sigma^{-1}\big] \big(\dot{\mathbf{x}}(t_2)-\mathbf{f}[\mathbf{x}(t_2)]\big), \] with \(\mathcal{C}(\tau)=\frac{\cot(\pi H)}{2\pi H}|\tau|^{-2H}\). This is the non-Markovian analogue of an Onsager–Machlup functional, and the same logic extends to the superdiffusive \(\ddot{\mathbf{x}}\)-based formulation [2209.11722]. ## 4. First-passage, wedge geometry, and visitation statistics The most developed first-passage problem for 2D fBm concerns wedge domains with absorbing boundaries. For a wedge \(0<\theta<\Theta\), exact Brownian results and exact FBM results at \(\Theta=\pi\) and \(\Theta=\pi/2\) motivate the long-time asymptotics \[ S_\Theta(t)\simeq t^{\pi(\alpha-2)/(2\Theta)} = t^{(\pi/\Theta)(H-1)}, \qquad \wp_\Theta(t)\simeq t^{-1+\pi(\alpha-2)/(2\Theta)}, \] where \(S_\Theta(t)\) is the survival probability and \(\wp_\Theta(t)\) the first-passage density. The case \(\Theta=\pi\) reduces to Molchan’s one-dimensional result, \[ S_\pi(t)\sim t^{H-1},\qquad \wp_\pi(t)\sim t^{H-2}, \] while \(\Theta=\pi/2\) yields \[ S_{\pi/2}(t)\sim t^{\alpha-2},\qquad \wp_{\pi/2}(t)\sim t^{\alpha-3}. \] In contrast, the method of images fails for FBM and gives the wrong one-dimensional exponent except at \(\alpha=1\) [1102.3633][1306.1667]. These wedge exponents imply a critical opening angle for the mean first-passage time, \[ \Theta_c=\pi(1-H)=\pi\left(1-\frac{\alpha}{2}\right). \] For \(\Theta<\Theta_c\) the mean escape time is finite; for \(\Theta\ge \Theta_c\) it diverges. Subdiffusive motion therefore admits finite mean escape even in relatively wide wedges, whereas persistent superdiffusive motion can have divergent mean escape times even in comparatively narrow wedges [1102.3633][1306.1667]. A complementary line of work studies territory coverage on a square lattice. The central result is that the inter-visit statistics of distinct sites are controlled by \(\mu=2H\). For recurrent and marginal 2D fBm (\(H\le 1/2\)), there is an early-time regime in which \[ F_n(\tau)\propto \tau^{-1-2H}, \] with no leading \(n\)-dependence in the prefactor. Characteristic times separate recurrent, marginal, and transient cases: \[ \vartheta_n=n^{1/(2H)},\quad \Theta_n\sim n^{1/(2H)} \qquad (H<1/2), \] \[ \vartheta_n=n^{1/2},\quad \Theta_n=n^{3/2} \qquad (H=1/2), \] \[ \vartheta_n=1,\quad \Theta_n=n^{1+1/(2H)} \qquad (H>1/2). \] For \(\vartheta_n\ll \tau\ll \Theta_n\), the survival tail of the inter-visit time obeys the stretched-exponential law \[ S_n(\tau)\sim \exp\!\left[-\left(\frac{\tau}{\vartheta_n}\right)^{\mu/(1+\mu)}\right], \] and for \(\tau\gg \Theta_n\) it crosses over to a simple exponential tail. Numerical simulations in 2D confirm the early power law, the stretched-exponential intermediate regime where present, and the predicted scaling of the crossover time [2407.11655]. These geometric laws map directly to reaction problems. In particular, one-dimensional three-particle survival problems can be rewritten as wedge problems for an effective 2D FBM, with wedge angle \(\Theta=\pi/3\) for the middle particle and \(\Theta=2\pi/3\) for an end particle. This produces exact asymptotic exponents for the corresponding reaction-survival probabilities and makes explicit how persistence and antipersistence alter encounter kinetics [1102.3633][1306.1667]. ## 5. Lévy area, rough paths, and stochastic integration The Lévy area is the antisymmetric second iterated integral of a planar path. For a 2D fBm \(B_t=(B_t^1,B_t^2)\), it is formally \[ A_t=\frac12\int_0^t \big(B_s^1\,dB_s^2-B_s^2\,dB_s^1\big). \] For \(1/31/2\), if \[ X_t=\int_0^t F(B_r)\,dB_r,\qquad Y_u=\int_0^u G(B_r)\,dB_r, \] then in two dimensions the covariance decomposes as \[ \mathbb{E}[X_tY_u] = \alpha_H\int_0^t\!\!\int_0^u \mathbb{E}\big(F(B_\tau)\cdot G(B_\sigma)\big) |\tau-\sigma|^{2H-2}\,d\sigma\,d\tau + \sum_{i,j=1}^2\int_0^t\!\!\int_0^u \gamma(\tau,\sigma)\, \mathbb{E}\big(\partial_iF_j(B_\tau)\partial_jG_i(B_\sigma)\big) \,d\sigma\,d\tau, \] where \(\alpha_H=H(2H-1)\) and \(\gamma(\tau,\sigma)\) is an explicit positive symmetric kernel. This formula shows that nonlinear integral functionals of 2D fBm are generally not themselves fractional Brownian motions. In particular, the stochastic integral component in the decomposition of the fractional Bessel process is not an fBm [1008.2500]. ## 6. Dependent components, spectral structure, and related extensions Dependent-component 2D fBm can be analyzed directly in both time and frequency domains. In the causal and well-balanced constructions with diagonal Hurst operator \(H=\operatorname{diag}(H_1,H_2)\), the increment process remains stationary, and the covariance of increments over lag \(h=t-s\) and step \(\delta\) is \[ \gamma^\Delta_{jk}(t,s) = \frac{\sigma_j\sigma_k}{2}\Big[ w_{jk}(h+\delta)|h+\delta|^{H_j+H_k} +w_{jk}(h-\delta)|h-\delta|^{H_j+H_k} -2w_{jk}(h)|h|^{H_j+H_k} \Big]. \] The low-frequency increment spectrum behaves as \[ \langle S^\Delta_{jk}(f)\rangle \sim \tilde c_{jk}\, f^{-(H_j+H_k-1)}, \qquad f\to 0. \] Hence \(H_j+H_k>1\) gives low-frequency divergence and long-range dependence, while \(H_j+H_k<1\) gives vanishing low-frequency cross-spectrum. In the well-balanced model the cross-spectrum is real; in the causal model it generally has a nonzero imaginary part, reflecting time-asymmetric dependence through \(\eta_{12}\neq 0\). A notable structural feature is that the construction is valid for the full parameter range \[ H_1,H_2\in(0,1),\qquad \sigma_1,\sigma_2>0,\qquad \rho\in[-1,1], \] without additional covariance-consistency constraints [2509.07537]. Simulation and inference follow this structure. A multivariate extension of the Wood–Chan circulant embedding algorithm generates stationary increment sequences with the required block Toeplitz covariance, and cumulative summation then produces 2D fBm sample paths. The same formulas suggest separate estimation of \(H_1,H_2\) from the marginal components and extraction of dependence from cross-covariance or cross-spectral quantities, with the imaginary part of the cross-spectrum distinguishing the causal from the well-balanced model [2509.07537]. A related, but distinct, extension is two-dimensional tempered fractional Brownian motion. Its probability density \(u(x,y,t)\) solves \[ \frac{\partial u}{\partial t} = D(t)\left(\frac{\partial^2u}{\partial x^2}+\frac{\partial^2u}{\partial y^2}\right), \] with \[ D(t)= \frac{\Gamma\!\left(H+\tfrac12\right)}{\sqrt{\pi}(2\lambda)^H} \,\lambda\, t^H K_{H-1}(\lambda t), \qquad 00. \] In the limit \(\lambda\to 0\), this recovers the standard 2D fBm Fokker–Planck equation \[ \frac{\partial u}{\partial t}=c_H t^{2H-1}\Delta_{x,y}u, \] with mean-square displacement \(\langle r^2(t)\rangle\propto t^{2H}\). For \(0