Model-Induced Stochasticity: Mechanisms & Dynamics

• Model-induced stochasticity is a phenomenon where inherent randomness is generated by a model’s internal mechanisms rather than by external noise.
• It is applied across fields such as gene regulatory networks, fluid dynamics, and neural network explanations to capture uncertainty and complex system transitions.
• This approach enhances models by revealing structural limitations and predicting phenomena like phase transitions, spontaneous stochasticity, and stability alterations.

Model-induced stochasticity denotes a class of phenomena in which randomness is generated, exposed, or parameterized by the structure of a model itself rather than treated solely as an exogenous perturbation. In the literature represented here, this includes stochasticity introduced at the level of update rules in gene regulatory networks, stochastic operators appended to inadequate closure models, random durations in operator splitting, stochastic regularisations of singular deterministic flows, weight-space perturbations in neural network explanations, and effective stochastic generators arising after conditioning on rare events (Murrugarra et al., 2012, Portone et al., 2017, Agazzi et al., 2022, Barlet et al., 11 Feb 2025, Bykov et al., 2021, Ghadermazi, 2019). Taken together, these works suggest that the term covers both explicit stochastic model components and limiting randomness that emerges when deterministic dynamics are coarse-grained, regularised, or rendered non-unique.

1. Conceptual scope

A recurring theme is that stochasticity can be introduced at the level of the model’s internal mechanisms rather than by appending a generic noise term to observables. In discrete gene regulatory networks, stochasticity is modeled at the biological function level: even if the expression levels of the input nodes of an update rule guarantee activation or degradation, there is a probability that the process will not occur due to stochastic effects (Murrugarra et al., 2012). In one-step models, the same idea appears as “stochastization,” where the dynamics are constructed from discrete random transitions using either a combinatorial formalism based on state vectors or an operator formalism based on occupation numbers (Kulyabov et al., 2019).

A second strand treats stochasticity as a representation of model inadequacy. In contaminant transport, a flawed embedded closure model is enriched with a stochastic error representation by introducing a stochastic operator, so that model-form uncertainty is propagated through the governing equations rather than absorbed into an error term on outputs (Portone et al., 2017). A related but distinct construction appears in the pathway model, where a pathway parameter switches the probability density among generalized type-1 beta, generalized type-2 beta, and generalized gamma forms, thereby producing thicker or thinner tails and changing the stochastic character of the modeled variable (Mathai et al., 2014).

A third strand concerns stochasticity that is revealed by effective descriptions of deterministic or locally interacting systems. In conditioned lattice gases, the auxiliary stochastic process that generates the biased ensemble of rare events generally has non-local dynamical rules, yet under certain constraints the effective process can remain local while becoming site-dependent (Ghadermazi, 2019). In rough flows and nearly-elastic systems, deterministic dynamics can exhibit limiting randomness once uniqueness fails or instabilities accumulate over long times, yielding what those works describe as spontaneous stochasticity or intrinsic stochasticity of the long-time behavior (Barlet et al., 11 Feb 2025, Freidlin et al., 2012).

2. Mechanisms by which models generate stochasticity

The mechanisms differ substantially across domains, but they are structurally comparable: each inserts randomness at a specific modeling layer.

Setting	How stochasticity enters	Representative paper
Rare-event ensembles on lattices	Effective generator after tilting and Doob transform	(Ghadermazi, 2019)
Deterministic vector-field dynamics	Sequentially following each vector field for a random amount of time	(Agazzi et al., 2022)
Deterministic black-box models	Random inputs chosen so outputs have prescribed statistics	(Mahmutoglu et al., 2012)
Neural network explanations	Multiplicative Gaussian noise in model weights	(Bykov et al., 2021)
Rough Lagrangian flows	Vanishing-noise or vanishing-uncertainty regularisation	(Barlet et al., 11 Feb 2025)
Discrete regulatory networks	Activation and degradation propensities within update rules	(Murrugarra et al., 2012)

In conditioned stochastic particle systems, the original generator is local and site-independent, but conditioning on rare events modifies the dynamics through a tilted generator and a Doob transform. The resulting effective dynamics are typically non-local or at least site-dependent because the eigenvector entering the transform encodes global information, although special constraints can preserve locality while inducing site dependence (Ghadermazi, 2019).

In random splitting of fluid models, stochasticity is injected by decomposing a deterministic vector field into minimal vector fields and evolving along each one for an independent, exponentially distributed random time of mean $h$. The overall process is Markovian and preserves structural features of the original conservative dynamics while converging almost surely to the deterministic model in the small noise limit (Agazzi et al., 2022).

In random input sampling, the model $h$ is deterministic, and the output randomness is induced entirely by the choice of random inputs. Because the desired output law generally determines the input distribution only implicitly, a modification of Metropolis-Hastings is used with a correction involving the density induced by a uniform input (Mahmutoglu et al., 2012). In NoiseGrad, the same structural idea is transferred to machine learning explanations: stochasticity is introduced directly into the weight parameter space, perturbing the decision boundary rather than the input (Bykov et al., 2021).

In spontaneous stochasticity for the WABC flow, the deterministic velocity field is only Hölder continuous for $h<1$, so existence and uniqueness fail in the limiting initial-value problem. Two stochastic regularisations are used numerically—Langevin-WABC and Cauchy-WABC—and both converge to the same nontrivial limiting distributions of trajectories (Barlet et al., 11 Feb 2025). This places the stochasticity in the regularisation procedure, but the limiting randomness is attributed to the singular deterministic dynamics themselves.

3. Mathematical representations

Several mathematical formalisms recur across the literature. For lattice particle systems, the stochastic generator is written as

$H = \sum_{i=1}^{L-1} h_{i,i+1},$

with a tilted generator

$H(s) = \sum_{i=1}^{L-1} h_{i,i+1}(s),$

and an effective generator

$H^{\mathrm{eff}} = V^{-1} H(s) V - \Lambda(s) I.$

Here the left eigenvector $V$ and leading eigenvalue $\Lambda(s)$ encode the conditioning on rare events; site dependence in the effective dynamics is tied to the structure of $V$ (Ghadermazi, 2019).

In the stochastic homeostasis model, the state evolves by the Markovian stochastic map

$a_{n+1} = s_n(\alpha a_n + \Delta),$

where the only source of noise is the multiplicative term $h$0 drawn from a distribution $h$1 with unit mean. This model yields distinct regimes separated by

$h$2

with a stationary heavy-tailed phase for $h$3 and total breakdown for $h$4 (Biswas et al., 2024).

In random splitting, the deterministic dynamics

$h$5

are replaced by randomized compositions of the flows of the $h$6. The transition kernel is

$h$7

and the process converges to the deterministic flow as $h$8 with error bounds of order $h$9 on bounded time intervals (Agazzi et al., 2022).

In discrete gene regulatory networks, stochasticity is specified by activation and degradation propensities. For node $h<1$0 with update function $h<1$1, the transition probability to the updated value is

$h<1$2

and joint transition probabilities factor as

$h<1$3

This produces a weighted state transition graph without requiring kinetic rates (Murrugarra et al., 2012).

In stochastic operator formulations for model inadequacy, the inadequacy enters as

$h<1$4

with a Fourier representation

$h<1$5

The stochasticity is then carried by the operator spectrum, often parameterized through the eigenvalues $h<1$6 or the full composite diffusion operator (Portone et al., 2017).

4. Dynamical consequences

The consequences of model-induced stochasticity are often qualitative changes in phase structure, stability, or observability that deterministic counterparts do not exhibit. In the homeostasis model, tuning the control parameter $h<1$7 produces a progression from homeostasis to a stationary heavy-tailed phase and then to catastrophic breakdown. The stationary distribution develops a power law tail

$h<1$8

with $h<1$9 determined by $H = \sum_{i=1}^{L-1} h_{i,i+1},$0 (Biswas et al., 2024). The same work distinguishes a “conspiracy principle” regime from a “catastrophe principle” regime and identifies reverse monotonicity of conditional exceedance as a signature of the catastrophic phase.

In ecological and evolutionary models, stochasticity can stabilize or destabilize depending on the balance between mean growth and variance. A central criticism of earlier persistence analyses is that invasion probabilities and times to extinction are not single-valued functions of $H = \sum_{i=1}^{L-1} h_{i,i+1},$1; when the diffusion approximation holds, these quantities are monotonic functions of $H = \sum_{i=1}^{L-1} h_{i,i+1},$2, with the threshold for stochasticity-induced stabilization given by $H = \sum_{i=1}^{L-1} h_{i,i+1},$3 (Dean et al., 2020). In a sexual speciation model, demographic stochasticity can generate phenotypic clusters in parameter ranges where deterministic models lead to a homogeneous distribution, and in some cases the effect is sizeable enough that deterministic modeling is insufficient (Lafuerza et al., 2015).

Long-time ecological dynamics can also be restructured. In community assembly models, mutualistic communities show little dependence on stochastic population fluctuations, whereas predator-prey models show strong dependence: the noise causes drastic decreases in diversity and total population size, and the power spectra shift from approximate $H = \sum_{i=1}^{L-1} h_{i,i+1},$4 without demographic noise to approximate $H = \sum_{i=1}^{L-1} h_{i,i+1},$5 with noise (Murase et al., 2010). In a slow-fast predator-prey system near a supercritical singular Hopf bifurcation, Gaussian white noise generates mixed-mode oscillations and recurrent outbreaks; the number $H = \sum_{i=1}^{L-1} h_{i,i+1},$6 of small oscillations between successive spikes is asymptotically geometric, with parameter related to the principal eigenvalue of a substochastic Markov chain (Sadhu, 2017).

Traffic modeling provides a distinct but related example. The stochastic Newell car-following model incorporates bounded acceleration, speed-dependent stochastic deceleration, and a larger randomization probability in the jam state. These modifications reproduce synchronized traffic flow and the concave growth pattern of oscillations that the deterministic Newell model cannot capture (Tian et al., 2016).

5. Biological, physical, and engineered realizations

Biological collective behavior provides a clear instance in which stochasticity is built into agent response functions. In a mathematical model of garden ants on a hexagonal lattice, each ant’s probability of choosing among neighboring sites is modulated by an individual stochasticity parameter $H = \sum_{i=1}^{L-1} h_{i,i+1},$7 through

$H = \sum_{i=1}^{L-1} h_{i,i+1},$8

The colony is modeled as a mixture of “normal ants” and “stochastic ants,” and the optimal distribution of stochasticity shifts with the feeding environment, yielding what the paper terms “diverse-stochasticity-induced optimization” (Shiraishi et al., 2018).

In memristive systems, stochasticity is tied to memory and nonlinear internal dynamics. A double-well memristor model uses

$H = \sum_{i=1}^{L-1} h_{i,i+1},$9

while a linear stochastic resonance model in a harmonic well adds correlated multiplicative and additive Gaussian white noises:

$H(s) = \sum_{i=1}^{L-1} h_{i,i+1}(s),$0

These models produce dynamic hysteresis, stochastic resonance, correlated-noise effects, and dirty hysteretic rounding (Gora et al., 2023).

In IMT neurons based on Vanadium Dioxide devices, spontaneous stochastic spiking is attributed to thermal noise and threshold fluctuations near bifurcation. The insulating-phase dynamics are modeled as an Ornstein-Uhlenbeck process with a fluctuating boundary,

$H(s) = \sum_{i=1}^{L-1} h_{i,i+1}(s),$1

and the firing rate is the inverse mean first-passage time. The experimentally observed firing probability is electrically controllable and exhibits a sigmoid-like transfer function (Parihar et al., 2017).

Noise-induced tipping point cascades provide another engineered reduction of stochastic dynamics. Starting from a chemically motivated individual-based model with Allee effects and dispersal, the dynamics are reduced to a four-state continuous-time Markov chain with states HH, HL, LH, and LL. Transition rates are reciprocals of mean first passage times, and the reduced model quantifies rescue effects and cascading collapse (Mallela et al., 2020).

6. Interpretation, invariance, and limitations

A common misconception is that stochasticity in modeling is synonymous with additive external noise. The surveyed literature shows a broader landscape: stochasticity may be attached to update propensities, closure operators, occupation-number formalisms, random splitting times, fluctuating boundaries, or regularisations of ill-posed deterministic equations (Murrugarra et al., 2012, Portone et al., 2017, Kulyabov et al., 2019, Agazzi et al., 2022, Parihar et al., 2017). In several cases, the point is not merely to represent ignorance but to capture mathematically unavoidable randomness in effective or limiting dynamics.

Another misconception is that induced stochasticity is necessarily an artifact of a chosen noise model. Two lines of work argue otherwise. In nearly-elastic systems with small energy loss at each collision, instabilities in the purely deterministic system lead to stochasticity of the long-time behavior, and the paper states that this stochasticity is an intrinsic property of the deterministic systems (Freidlin et al., 2012). In the WABC model, the observed spontaneous stochasticity does not depend on the chosen stochastic regularisations: Langevin-WABC and Cauchy-WABC converge to the same nontrivial limiting distributions within numerical precision (Barlet et al., 11 Feb 2025).

At the same time, the literature emphasizes constraints and diagnostics. In effective rare-event dynamics, locality of the effective generator is exceptional and requires specific algebraic constraints on the microscopic reaction rules (Ghadermazi, 2019). In stochastic operator approaches to model inadequacy, calibration with insufficiently informative data can fit noise and generate nonphysical solutions, so physical constraints on the operator and sufficiently rich data are essential (Portone et al., 2017). In ecological persistence theory, the choice of metric matters: explicit formulas for invasion probabilities and persistence times are monotonic functions of $H(s) = \sum_{i=1}^{L-1} h_{i,i+1}(s),$2 under the diffusion approximation, whereas they are not single-valued functions of $H(s) = \sum_{i=1}^{L-1} h_{i,i+1}(s),$3 (Dean et al., 2020).

Taken together, these results support a precise interpretation of model-induced stochasticity: it is not a single mechanism but a family of mathematically explicit constructions by which modeling choices, structural reductions, and singular limits generate random dynamics, random effective parameters, or random ensembles with measurable dynamical consequences.