Rank-Based Interacting Systems

• Rank-based interacting systems are dynamical models where agent interactions depend on their ordered statistics, introducing discontinuities and non-exchangeable dynamics.
• They are applied in various fields such as the Atlas model in finance, network ranking, and branching processes, illustrating practical use in modeling equity markets and complex systems.
• Analytical methods including mean-field limits, Laplace transforms, and statistical inference techniques rigorously characterize stationary measures, chaoticity, and hydrodynamic limits.

A rank-based interacting system is a class of dynamical model in which each agent or particle receives interactions, drift, or other coefficients determined explicitly by its current rank or ordering within the system—often via its position among the ordered statistics of all agents. This framework underpins much of modern stochastic modeling in areas ranging from mathematical finance (the Atlas model, stochastic portfolio theory), to nonlinear diffusions, branching systems, network ranking, and even finite dynamical systems and interacting agent models. Rank-based interactions introduce discontinuities at collision manifolds, non-exchangeable dependencies, and often yield rich mean-field and hydrodynamic limits.

1. Mathematical Formulation and Mean-Field Limits

Given $n$ particles, the canonical rank-based system assigns prescribed state-dependent drift and possibly volatility to each particle indexed by its rank at time $t$. Denote the ordered particle values by $(X_{(1)}(t)\le ... \le X_{(n)}(t))$ and let rank function $\operatorname{rank}(X_i(t)) = k$ iff $X_i(t) = X_{(k)}(t)$.

Typical finite-$n$ dynamics: $$dX_i(t) = b_n(\operatorname{rank}(X_i(t)))\,dt + \sigma\,dW_i(t)$$ where $b_n(k)$ is derived from an underlying function $b:[0,1]\to\mathbb{R}$ by

$b_n(k) = n \int_{(k-1)/n}^{k/n} b(v)\,dv$

and $\sigma>0$ is volatility.

In the mean-field limit $n\to\infty$, the empirical distribution $\mu_t$ (law of a generic particle) has cumulative distribution $F_t(x) = \mu_t((-\infty,x])$, and the limiting McKean–Vlasov SDE reads

$dX(t) = b(F_t(X(t)))\,dt + \sigma\,dW(t)$

This nonlinear SDE corresponds to a deterministic evolution of the population law governed by a generalized porous medium equation: $$\partial_t G = -\partial_x B(G) + \frac{1}{2}\partial_x^2 E(G)$$ where $B(r) = \int_0^r b(a)\,da$, $E(r) = \int_0^r \sigma^2(a)\,da$ (Shkolnikov, 2010, Reygner, 2014, Kolli et al., 2016). These equations admit unique, well-posed solutions under mild regularity and monotonicity conditions.

2. Stationary Measures, Chaoticity, and Laplace Transforms

In regimes with decreasing drift $b$ and zero-sum property ($B(1)=0$), the stationary density for the mean-field limit is

$p_\infty(x) = \frac{2 B(F_\infty(x))}{\sigma^2}$

with quantile function

$$\Phi(u) = \int_0^u \frac{v}{2B(v)}\,dv - \int_u^1 \frac{1-v}{2B(v)}\,dv$$

so that $F_\infty^{-1}(u) = \Phi(u)$.

The Laplace transform is

$L_\infty(r) = \int_{u=0}^1 e^{r \Phi(u)}\,du$

For finite-$n$, the (projected) stationary law has density proportional to $$\exp\big(\frac{1}{\sigma^2} \sum_{k=1}^n b_n(k) z_{(k)}\big)$$. Explicit product formulas for two-point Laplace transforms enable rigorous proof of “chaoticity”—convergence of finite-dimensional marginals (in all Wasserstein orders) to independent copies of the stationary law as $n\to\infty$ (Reygner, 2014).

3. Key Applications: Atlas Model, Zipf's Law, and Equity Markets

3.1 Atlas Model

In the Atlas model, particles (interpreted as stock capitalizations) have drift

$$d\log X_i(t) = \left(-g + n g \mathbf{1}_{\{r_t(i) = n\}} \right) dt + \sigma\,dW_i(t)$$

where only the lowest-ranked stock (“Atlas”) receives a compensating upward drift. The stationary distribution of normalized capital weights yields a Pareto law; specifically,

$$\theta_{(k)} \propto k^{-\alpha}\quad\text{with}\quad \alpha = \frac{\sigma^2}{2g}$$

Zipf’s law ($\alpha=1$) arises precisely when $\sigma^2 = 2g$ (Fernholz et al., 2016).

3.2 Stochastic Portfolio Theory

General linear and piecewise rank-based drifts model equity markets; market-stability results show that as $n\to\infty$, the stationary curve converges to a deterministic profile, and empirically computed growth/turnover rates converge to those from the nonlinear diffusion equilibrium (Reygner, 2014, Shkolnikov, 2010).

3.3 Branching and Go-or-Grow Models

Branching rank-based models (e.g., “Go-or-Grow” systems: only $K$ highest-ranked particles branch, others receive drift) yield hydrodynamic limits in the form of free-boundary PDEs with discontinuous coefficients. These systems manifest pulled and pushed traveling waves, Bramson correction effects, and transition thresholds for propagation speed (Demircigil et al., 13 May 2025).

4. Extensions: State-Dependence, Common Noise, and Generalizations

The basic Atlas/rank-based SDE can be extended to allow state-dependent drift and volatility: $$dX^i_t = \sum_{k=1}^N b_k(X^i_t) \mathbf{1}_{X^i_t = X^{(k)}_t} dt + \sum_{k=1}^N \sigma_k(X^i_t) \mathbf{1}_{X^i_t = X^{(k)}_t} dW^i_t$$ leading to degenerate, discontinuous dynamics novel in both finance (e.g., size-dependent volatility in fish-pond or population models) and statistical physics. Recent work establishes strong well-posedness in $\mathbb{R}^2$, weak well-posedness in higher dimensions, and positivity under vanishing diffusion at boundary for appropriate drift choices (Guérin et al., 10 Jan 2026).

With common noise, the mean-field limit becomes stochastic—empirical CDF evolves under a rough stochastic flux, leading to SPDEs akin to stochastic conservation laws: $$dG(t,x) = [ -\partial_x B(G) + \partial_{xx} E(G) + \frac{1}{2} y^2 \partial_{xx} G ] dt - y \partial_x G dW(t)$$ This formulation connects with recent theory in rough-paths and stochastic conservation law regimes (Kolli et al., 2018).

In finite dynamical systems (over finite alphabets), “rank” refers to the cardinality of the image; maximum rank and periodic rank are determined by underlying interaction graph geometry (via the $\alpha_p(D)$ invariant), and update schedule (parallel, block-sequential, complete). Rank-based design controls reversibility and attractor structure (Gadouleau, 2015).

5. Rank-Based Systems in Network Science and Machine Learning

Rank-based mechanisms drive network inference and ranking tasks:

• SpringRank: Assigns a real-valued hierarchy $r_i$ to nodes; edges are modeled as “springs,” favoring interactions between nodes of similar rank. Efficient quadratic minimization yields ranks; the method supports edge prediction, significance testing, and scales to large graphs (Bacco et al., 2017).
• Mixed Model (XOR Rank-Community): Combines mixed-membership SBM and spring-energy ranking; each node chooses between community or rank-driven interaction, captured via latent $\sigma_i \in \{0,1\}$, with probabilistic EM inference and per-node assignment (Iacovissi et al., 2021).
• Rank-Based Supplanting Process: Models macaque-like dominance with supplanting dynamics and emergent rank-correlated centrality; overlap centrality $C_i$ becomes perfectly correlated with agent rank as supplanting rate $p \to 0$ (modulo singularities) (Shimomura et al., 2022).
• Network Reconstruction and Online Rankers: In protein interaction prediction, coordinated online ranking models assign local and global rankers, regularized by neighborhood geometry, improving candidate prioritization in experimental settings (Bar-Shira et al., 2011).

6. Statistical Estimation and Model Inference for Rank Data

In social and behavioral sciences, exponential-family random graph models (ERGMs) handle rank-order data (each ego ranks all alters), with statistics built from pairwise ordinal comparisons. Sufficient statistics encode exogenous covariates, homophily, non-conformity, deference aversion, and dynamic changes (inertia), estimated by MCMC maximum likelihood (Krivitsky et al., 2012).

7. Integrable Systems and Rank-Reduction Phenomena

Rank (in the sense of matrix or group rank) plays a central role in reduction of complex, high-dimensional integrable models. Explicit reductions of rank-2 Hitchin systems on genus $g=2,3$ curves yield finite-dimensional interacting particle systems; universality of the reduction is demonstrated by Lagrange interpolation polynomials, and Liouville integrability established via commuting Hamiltonians (Sheinman, 2017).

---

Rank-based interacting systems constitute a multi-disciplinary paradigm, bridging stochastic processes, statistical physics, financial modeling, network science, and integrable systems. Mathematical analysis centers on the impact of ordering-induced discontinuities, nontrivial limiting PDEs, and universal statistical properties derived from ranking logic. Convergence, stationarity, chaos, and the role of rank in dynamics and inference are central technical themes across the literature.