Planar System with Rank-Dependent Drift

• Planar systems with rank-dependent drift are defined by stochastic processes whose drift and diffusion coefficients switch based on the relative ordering of their coordinates.
• The analysis employs piecewise-defined stochastic differential equations, local time arguments, and skew Brownian motion representations to handle collisions and discontinuities.
• These systems have practical applications in stochastic portfolio theory, statistical physics, and interacting particle models, ensuring strong existence and uniqueness under certain conditions.

A planar system with rank-dependent drift consists of two interacting stochastic processes on $\mathbb{R}^2$, driven by Brownian motions, whose drift and diffusion coefficients at any time depend only on the current order (rank) of their coordinates rather than on their labels. Rank-based dynamics emerge prominently in stochastic portfolio theory, statistical physics, and the theory of interacting particle systems. Such systems exhibit degenerate, piecewise-continuous coefficients at the collision manifold $x_1=x_2$, and their analysis returns to classical results on local times, strong uniqueness, skew Brownian motion, and generalized Tanaka-type stochastic differential equations (SDEs).

1. Formulation of the Rank-Dependent Planar SDE

The canonical planar rank-dependent system concerns two processes $X_1(t), X_2(t)$ with SDEs: $$\begin{aligned} dX_1(t) &= [g \, 1_{X_1 \le X_2} - h \, 1_{X_1 > X_2}] \, dt + [\rho \, 1_{X_1 > X_2} + \sigma \, 1_{X_1 \le X_2}] \, dB_1(t),\ dX_2(t) &= [g \, 1_{X_1 > X_2} - h \, 1_{X_1 \le X_2}] \, dt + [\rho \, 1_{X_1 \le X_2} + \sigma \, 1_{X_1 > X_2}] \, dB_2(t) \end{aligned}$$ with $\rho^2+\sigma^2=1$, $g,h \geq 0$, $B_1,B_2$ independent standard Brownian motions (Fernholz et al., 2011). Drift and dispersion switch according to the ordering of coordinates—at any instant, the "leader" and "laggard" are assigned $(g,-h), (\sigma, \rho)$ or vice versa.

Extensions to general state-dependent drift and non-uniform diffusion are studied with locally Lipschitz $b_1,b_2: \mathbb{R} \to \mathbb{R}$ and $\sigma_1,\sigma_2: \mathbb{R} \to [0,\infty)$ (Guérin et al., 10 Jan 2026). For $x \in \mathbb{R}^2$: $$b^1(x) = \begin{cases} b_1(x_1) & x_1 < x_2, \ b_2(x_1) & x_1 > x_2, \end{cases} \quad b^2(x) = \begin{cases} b_1(x_2) & x_2 < x_1, \ b_2(x_2) & x_2 > x_1, \end{cases}$$ with diffusion matrix $\operatorname{diag}(\sigma_1(x_1), \sigma_2(x_2))$.

When restrictions to the nonnegative quadrant are imposed, normal reflection along axes and rank-dependent coefficients on the faces are incorporated, including local time terms to realize reflection (Ichiba et al., 2012).

2. Infinitesimal Generator and the Order Statistic Structure

The infinitesimal generator $L$ acts on $C^2$-functions $f(x_1,x_2)$ via: $$\begin{aligned} Lf(x) &= 1_{x_1 > x_2}\left[ \frac{\rho^2}{2}\frac{\partial^2 f}{\partial x_1^2} + \frac{\sigma^2}{2}\frac{\partial^2 f}{\partial x_2^2} - h \frac{\partial f}{\partial x_1} + g \frac{\partial f}{\partial x_2} \right]\ &+ 1_{x_1 \le x_2}\left[ \frac{\sigma^2}{2}\frac{\partial^2 f}{\partial x_1^2} + \frac{\rho^2}{2}\frac{\partial^2 f}{\partial x_2^2} + g \frac{\partial f}{\partial x_1} - h \frac{\partial f}{\partial x_2} \right] \end{aligned}$$ The generator depends solely on rank—(max, min)—and underlies the Markovian structure of the ranked process $(R_1, R_2) = (\max\{X_1, X_2\},\min\{X_1, X_2\})$ (Fernholz et al., 2011, Itkin et al., 2021). The ranked processes admit dynamics governed by local time at collisions: $$\begin{aligned} dR_1(t) &= \text{leader SDE} \; + \frac{1}{2} dL^{R_1-R_2}(t),\ dR_2(t) &= \text{laggard SDE} \; - \frac{1}{2} dL^{R_1-R_2}(t). \end{aligned}$$ Collision phenomena are handled via local time, which removes stickiness and resolves singularities at $x_1 = x_2$.

3. Transition Densities, Skew Brownian Representation, and Local Time

The difference $Y(t) = X_1(t) - X_2(t)$ in the canonical constant-coefficient case evolves as a skew Brownian motion with bang-bang drift: $$dY(t) = -\lambda \, \operatorname{sgn}(Y(t)) \, dt + dW(t), \quad \lambda = g + h,$$ where $W$ is a Brownian motion derived from the noise terms. The full joint law of $(X_1, X_2)$ is captured via the local time of $Y$ at $0$ and an independent Brownian summand, yielding explicit formulas for the transition density—especially in isotropic and degenerate regimes (Fernholz et al., 2011).

In systems with skew-elastic collision regimes, local times $L^Y(t)$ (right and left) at zero for $Y$ encode interaction patterns from frictionless crossing to perfect reflection. The rank-based SDE system can be reduced to a one-dimensional SDE for $Y$: $$Y(t) = y - \lambda \int_0^t \overline{\operatorname{sgn}}(Y(s)) ds + W(t) + 2(2\alpha-1)\widehat{L}^Y(t),$$ with $\alpha\in[0,1]$ parametrizing the collision regime (Fernholz et al., 2012). Transition densities are available in closed form, e.g. equation (4.13) in (Fernholz et al., 2012).

Skew Brownian representations of the full $(X_1,X_2)$ system are possible, with additional terms in the SDE driven by the local time along the diagonal (Fernholz et al., 2011, Fernholz et al., 2012).

4. Existence, Uniqueness, and Strong Well-Posedness

Strong existence and pathwise uniqueness up to possible explosion hold under general conditions. With constant coefficients and nondegenerate diffusion, pathwise uniqueness and strong solutions are established globally for $n=2$ due to the absence of triple collisions (Fernholz et al., 2011, Ichiba et al., 2011, Itkin et al., 2021). Arguments rely on:

• Localization away from the collision set $\Theta=\{x_1=x_2\}$, where the SDE is locally Lipschitz.
• The construction of $C^1$ diffeomorphisms that smooth out drift discontinuities near $\Theta$ (Guérin et al., 10 Jan 2026).
• The use of generalized Tanaka equations and strong uniqueness criteria for SDEs with bounded variation coefficients (Fernholz et al., 2011, Fernholz et al., 2012).

When reflecting boundary conditions are imposed (quadrant case), strong existence holds up to corner hitting, and is global when the laggard’s variance dominates (Ichiba et al., 2012). The non-coalescence and non-stickiness of collisions are established using occupation density arguments, ensuring instantaneous exit from diagonal collisions (Itkin et al., 2021).

5. Time Reversal and Singular Terms

Under time reversal, the process continues to satisfy an SDE of rank-dependent form, but the drift acquires an extra Nelson-type term involving the gradient of the transition kernel: $$d\widehat{Y}(t) = [\lambda \operatorname{sgn}(\widehat{Y}(t)) + \partial_\xi \log p_{T-t}(0, \widehat{Y}(t))] dt + dW^\sharp(t)$$ The local time terms persist, and their behavior under time-reversal is analyzed explicitly (Fernholz et al., 2011, Fernholz et al., 2012). In stationary cases, strict time-reversibility and bridge drift formulae are established.

6. Extensions, Generalizations, and Applications

Extensions to state-dependent coefficients, degenerate diffusions, and systems with reflecting boundaries (orthant-valued, Skorokhod reflection) have been constructed (Ichiba et al., 2012, Guérin et al., 10 Jan 2026). The Dirichlet form approach admits ergodicity proofs, strong Feller properties, and accommodates broad classes relevant in stochastic portfolio theory (Atlas, volatility-stabilized, polynomial models) (Itkin et al., 2021). The planar system serves as a model for complex interactions in higher-dimensional rank-based systems, for which collision structure and solvability hinge on detailed conditions involving concavity of volatility profiles and nondegeneracy.

These systems furnish canonical examples of interacting stochastic particles with discontinuous coefficients tied to relative position rather than absolute identity, and are of ongoing interest in mathematical finance, physics, and the theory of stochastic processes.