Stochastic IFS in Random Dynamics & Fractal Geometry

• Stochastic IFS are a class of random systems defined on metric spaces using probabilistically selected continuous maps to create fractal attractors.
• They guarantee unique invariant measures under contraction or state-dependent criteria, ensuring ergodicity and stability in the system.
• Applications include modeling random fractals, economic dynamics, and algorithmic graphics, where randomness enriches the attractor geometry.

A stochastic iterated function system (IFS) is a random dynamical system defined on a metric space, modeled as a finite or parameterized family of continuous maps applied in random order according to a probability law. This framework generalizes deterministic IFS by introducing randomness either through probabilistic selection of maps or, in the most general versions, through state-dependent or measure-dependent probability kernels. Stochastic IFS are primary models for random fractals, probabilistic self-similarity, and random dynamical attractors. Their invariant measures and attractors are central to modern probability, ergodic theory, fractal geometry, random dynamical systems, and applications across mathematics, physics, engineering, and economics.

1. Foundational Definitions and Notation

Let $(X,d)$ be a complete (typically compact or bounded) metric space. Classical stochastic IFS (also called iterated random function systems) consist of:

• A collection of continuous (often $C^1$) maps $f_i:X\to X$, $i=1,\dots,N$.
• Probabilities $p_i>0$ with $\sum_i p_i=1$ for the choice of $f_i$ at each iteration.
• An i.i.d. sequence of random indices $\{\omega_n\}$ drawn with $\mathbb{P}(\omega_n=i)=p_i$.

A stochastic orbit is given by $x_{n+1}=f_{\omega_n}(x_n)$. The associated transfer operator $F: \mathcal{P}(X) \to \mathcal{P}(X)$ on probability measures is

$$F(\mu) = \sum_{i=1}^N p_i \, {f_i}_\# \mu, \qquad ({f_i}_\# \mu)(A) = \mu(f_i^{-1}(A)).$$

The invariant measure $\mu^*$ satisfies $F(\mu^*) = \mu^*$, and its support defines the fractal attractor.

Generalizations include place-dependent probabilities $p_i(x)$, continuous parameterizations $\lambda \in \Lambda$, and IFSm (iterated function systems with general measures), where the random choice is governed by a family $q_x$ of probability measures on the parameter space $\Lambda$ (Oliveira et al., 14 May 2025, Ghosh et al., 2022).

2. Existence, Uniqueness, and Stability of Invariant Measures

The existence and uniqueness of an invariant (Hutchinson) measure $\mu^*$ and its support as attractor depend on contractivity properties:

• If each map $f_i$ is a contraction with constant $s_i<1$, and the average contraction $L=\sum_i p_i s_i<1$, then $F$ is a strict contraction in the Wasserstein-1 metric, ensuring a unique $\mu^*$ (Bouke, 24 May 2025, Ghosh et al., 2022, Wang, 2012).
• For state-dependent probabilities $p_i(x)$ and Lipschitz maps $f_i$, similar contraction-in-expectation conditions guarantee unique invariant measures and geometric ergodicity, per theorems of Hutchinson–Barnsley and extensions by Tyran-Kaminska and others (Kungurtsev et al., 2021, Oliveira et al., 14 May 2025).

Stability and ergodicity can be established via Lyapunov-type (negative expected logarithmic derivative) or mean-square contraction criteria; e.g.,

$$\mathbb{E}[\log \|Df_{\omega}(x)\|] < 0,\quad \mathbb{E}[d(f_i(x),f_i(y))^2] \leq \lambda d(x,y)^2 \text{ with } \lambda<1.$$

Such conditions imply geometric ergodicity and almost-sure convergence of random orbits to the invariant law (Bouke, 24 May 2025, Kungurtsev et al., 2021). Even in non-hyperbolic systems, under “locally injective” and mild separability assumptions, stationary measures exist and are supported on generalized attractors (Matias et al., 2016).

3. Attractor Geometry, Dimension, and Computational Methods

The invariant measure $\mu^*$ is typically supported on a fractal attractor. For finite IFS: $$A = \bigcup_{i=1}^N f_i(A), \quad \text{where } A = \operatorname{supp} \mu^*.$$ Fractal dimension is commonly estimated by the box-counting method: $$D_0 = \lim_{\epsilon\to 0} \frac{\log N(\epsilon)}{-\log \epsilon},$$ where $N(\epsilon)$ is the minimal number of boxes of side $\epsilon$ covering the observed trajectory points (Bouke, 24 May 2025). For random nonlinear IFS, empirical studies have produced dimensions ranging from $\approx 1.40$ (for disruptive mixtures) up to $\approx 1.89$ (for strongly oscillatory maps), demonstrating that stochasticity and nonlinearity enrich internal attractor geometry well beyond classical deterministic cases (Bouke, 24 May 2025).

Computationally, the "chaos game" algorithm underlies both visualization and empirical study: random orbits are iterated, recorded (after burn-in), and used to approximate the measure and attractor (Ghosh et al., 2022, Bouke, 24 May 2025, Barnsley et al., 2012). See the table for algorithmic pseudocode details.

Algorithm Step	Description	Source
Initialize $x_0$	Start at $x_0 \in X$	(Bouke, 24 May 2025)
For $n=1$ to $M$	Randomly pick $f_i$ per $p_i$ and set $x_{n}=f_i(x_{n-1})$	(Bouke, 24 May 2025)
Record $x_n$ (post burn-in)	Gather sample for attractor approximation	(Bouke, 24 May 2025)
Box-counting analysis	Estimate $D_0$ from point cloud	(Bouke, 24 May 2025)

4. Extensions: State-Dependent and Measure-Dependent Frameworks

State-dependent probabilities $p_i(x)$ and continuous parameter IFS with family $q_x$ have been formalized as IFSm (Iterated Function Systems with Measures) (Oliveira et al., 14 May 2025, Kungurtsev et al., 2021). In IFSm, for $(X,d)$ and compact $\Lambda$, the random map $\tau_\lambda(x)$ is applied with probability $q_x$ on $\Lambda$; the Markov operator acts as

$$T_q \mu(A) = \int_X q_x\left(\{\lambda : \tau_\lambda(x)\in A\}\right) d\mu(x).$$

General existence and uniqueness criteria employ uniform contraction, joint Lipschitz regularity, and continuity assumptions for $q_x$ and $\tau_\lambda(x)$, extending Banach’s fixed point argument to this infinite-dimensional setting. Under suitable conditions, the support of the invariant measure equals the attractor, and the system exhibits stochastic stability under perturbations of $q_x$ (Oliveira et al., 14 May 2025).

These generalized frameworks encompass models with place-dependent stochasticity, random parameterizations, and non-affine morphisms, increasing both expressivity and analytical complexity.

5. Nonlinear and Non-Hyperbolic Systems

Random nonlinear IFS (RNIFS) extend classical affine models, incorporating $C^1$ (or more regular) nonlinear maps and arbitrary probability laws. Recent advances have proven that under average contractivity conditions—e.g., Lyapunov negative exponents, mean-square contraction—there exist unique invariant measures, and orbits converge in law (Bouke, 24 May 2025).

Non-hyperbolic stochastic IFS can possess attractors and stationary measures even without uniform contractions, provided conditions such as weak hyperbolicity or local injectivity are met. For example, if all maps are injective on some interval and the attractor is separable, almost sure convergence results and the existence of non-atomic invariant measures are guaranteed (Matias et al., 2016). This broadens the class of systems to which stochastic IFS theory applies, including those with expanding or mixed-type constituents.

6. Applications and Empirical Studies

Stochastic IFS have a diverse span of applications:

• Fractal geometry and random fractals: Random $\beta$-transformations on the fat Sierpiński gasket, with rigorously determined invariant measures of maximal entropy and absolutely continuous invariant measures for suitable parameter regimes (Zhang et al., 2022).
• Economic modeling: Random utility and optimal stochastic growth models formulated as Markov chains generated by stochastic affine IFS, generating long-run distributions with fractal support (Wang, 2012).
• Stochastic control: Model predictive control under uncertainty represented as a stochastic IFS, where ergodicity and geometric convergence are derived from contraction-in-expectation and kernel continuity (Kungurtsev et al., 2021).
• Algorithmic graphics and simulation: The chaos game, visualization of fractal attractors for both deterministic and random cases (Ghosh et al., 2022, Barnsley et al., 2012).
• Statistical physics, finance, and biological modeling: Stochastic IFS underpin models of DNA replication, opinion dynamics, option pricing, and networked systems (Ghosh et al., 2022).

Empirical studies underscore the sensitivity of attractor geometry to balancing between nonlinearity and probabilistic weighting—demonstrated by increased fractal dimension when nonlinear, oscillatory maps are well-mixed (Bouke, 24 May 2025).

7. Open Challenges and Research Frontiers

Areas of ongoing research and open questions include:

• Non-asymptotic mixing rates and large deviations for empirical measures,
• Continuous-time and hybrid IFS,
• IFS with discontinuous or data-driven probabilities,
• Rigorous numerical approximation of invariant measures with error bounds,
• Set-valued or non-autonomous IFS, and
• General criteria for existence/uniqueness in data-rich or nonparametric settings (Ghosh et al., 2022, Oliveira et al., 14 May 2025).

Stochastic IFS frameworks, both in their classical and generalized forms, continue to expand in theoretical scope and application, with ongoing advances in analysis, simulation, and interdisciplinary deployment.