🌀 nonlinear sciences

Intermittency induced by long memory under stochastic regime switching

This paper establishes that the interaction between long-range memory and stochastic regime switching in nonlinear dynamical systems creates a fundamental annealed-quenched dichotomy, where systems can remain stable in expectation while exhibiting pathwise intermittent bursts with heavy-tailed statistics due to the amplification of rare supercritical excursions.

Original authors: Mauricio Herrera-Marín

Published 2026-05-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mauricio Herrera-Marín

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a complex system, like a busy city's traffic network or a school of fish, that is constantly changing its rules. Sometimes the rules are calm and orderly; other times, they are chaotic and prone to explosions of activity. This paper studies what happens when such a system has two specific traits: long memory and random rule-switching.

Here is the core story of the research, explained simply:

The Two Main Characters

The System with Long Memory: Think of this system like a person with a very long memory who never forgets a compliment or a criticism. If the system gets a little boost of energy today, that boost doesn't just disappear; it lingers, echoes, and gets reinjected into the system for a long time. In math terms, this is called "Volterra evolution" or "fractional memory."
The Random Switcher: Now, imagine the rules of the game change randomly. Sometimes the system is in a "Safe Zone" where things die down. Other times, it randomly jumps into a "Danger Zone" where things tend to grow explosively. This is modeled by a "Markov chain," which is just a fancy way of saying a random switcher that jumps between states.

The Big Surprise: The "Annealed" vs. "Quenched" Trap

The paper discovers a fascinating and dangerous disconnect between how the system looks on average versus how it actually behaves in real life.

The "Annealed" View (The Average): If you take a snapshot of the system over a long time and average out all the chaos, it looks perfectly safe. The math says, "On average, the system is stable. The explosions are rare enough that the average size of the system stays small." It's like saying, "On average, a gambler doesn't lose much money," because most days they win small amounts, and the big losses are rare.
The "Quenched" View (The Real Path): But if you watch a single specific run of the system (a single "sample path"), the story is very different. Because the system has long memory, if it accidentally wanders into the "Danger Zone" and stays there for a slightly longer time than usual, that memory acts like an amplifier. It takes that short period of chaos and stretches it out, turning a small spark into a massive, long-lasting fire.

The Metaphor:
Imagine a forest that usually stays calm (the average view). However, the trees have roots that hold onto moisture for weeks. If a lightning strike happens (the random switch to a dangerous state) and the fire burns for just a few extra minutes longer than usual, the moisture in the roots keeps the fire smoldering and spreading long after the lightning is gone.

The Average: "Most of the time, the forest is fine."
The Reality: "When the fire does start, it becomes a massive, uncontrollable inferno because the forest 'remembers' the heat."

The Mechanism: Why It Happens

The paper explains that this "intermittency" (the irregular bursts of activity) happens because of a specific combination:

Rare Events: The system occasionally stays in the "Danger Zone" for a long time.
Memory Amplification: The system's long memory takes that long stay and turns it into a huge burst of energy.

Even if the "Danger Zone" is visited rarely, the memory ensures that once it is visited, the damage is permanent and massive. This creates a situation where the system is stable in expectation (mathematically safe on average) but unstable in reality (prone to massive, heavy-tailed bursts).

The Micro-Macro Connection

The researchers also showed that this isn't just a theoretical math trick. They proved that if you have a huge crowd of individual agents (like neurons firing or people making trades) that are all reacting to each other and switching rules randomly, the entire crowd behaves exactly like the single "long-memory" system described above.

Micro: Individual agents branching out and reacting.
Macro: The whole group suddenly exploding in a coordinated burst.
The paper shows that the "burst" mechanism travels from the tiny individual level all the way up to the big picture level.

What the Numbers Show

The authors ran computer simulations to prove this. They found:

Bursts are Heavy-Tailed: The size of these explosions follows a "power law." This means while small bursts are common, massive, system-crashing bursts are much more likely than you would expect in a normal, random system.
Geometry Matters: If the system is a network (like a social network or a power grid), the shape of that network determines where the burst happens. Sometimes the energy focuses on one specific part of the network, and sometimes it spreads out, depending on how the "rules" interact with the network's shape.
Non-Commutativity: Even if the rules don't play nicely together (mathematically speaking, they don't "commute"), the bursts still happen. The system doesn't need perfect symmetry to explode; it just needs memory and random switching.

The Bottom Line

The paper warns us that looking at averages is dangerous when dealing with systems that have long memory and random changes.

If you only look at the average, you might think a system is safe.
But if you look at the actual path the system takes, you might see it building up to a massive, unpredictable explosion.

The authors call this the "Annealed-Quenched Dichotomy." It's a reminder that in complex, memory-rich systems, the "average" story can be a lie, and the real risk lies in the rare, long stretches of bad luck that the memory refuses to forget.

1. Problem Statement

The paper investigates a fundamental instability mechanism in nonlinear, nonlocal dynamical systems arising from the interaction of two specific features:

Long-range memory: Modeled via operator-valued Volterra evolution equations with completely monotone memory kernels (e.g., fractional kernels).
Stochastic regime switching: The system parameters (excitation operators and kernel weights) are modulated by an ergodic, finite-state continuous-time Markov chain.

The Core Question: How do stability, amplification, and intermittency emerge when these mechanisms interact? Specifically, the author addresses the paradox where a system may be stable in expectation (annealed stability) but exhibit pathwise growth and heavy-tailed bursts (quenched instability) due to rare, long residence times in supercritical regimes amplified by memory.

2. Mathematical Framework

The system is modeled by a randomly switching Volterra equation on an augmented state space $X = H \times H$ :
$U'(t) = B U(t) + \int_0^t G_{Z(t)}(t-s) U(s) \, ds + F_{Z(t)}$

$Z(t)$ : An ergodic continuous-time Markov chain on a finite state space $\mathcal{Z} = \{1, \dots, m\}$ .
$B$ : A dissipative linear operator.
$G_z(t)$ : An operator-valued memory kernel with a structured form involving a bounded excitation operator $A_z$ and a scalar completely monotone kernel $g_z(t)$ .
Non-Markovian Nature: Because $U(t)$ depends on the full history via the convolution, the joint process $(U(t), Z(t))$ is generally non-Markovian in finite dimensions. The state must be augmented (e.g., via Bernstein representation) to restore Markovian properties.

3. Methodology

The paper employs a multi-faceted approach combining functional analysis, probability theory, and numerical simulation:

Pathwise Well-posedness: Establishes existence and uniqueness of mild solutions for fixed sample paths of the regime process using a Picard-Volterra fixed-point argument.
Annealed Analysis (Expectation-based):
- Derives Volterra-type a priori bounds to control moments.
- Constructs a memory-adapted Lyapunov functional using the Bernstein representation of completely monotone kernels. This allows for mean-square dissipation estimates under an averaged subcriticality condition.
Quenched Analysis (Pathwise):
- Analyzes the amplification of rare events where the system stays in a supercritical (unstable) regime for an unusually long time.
- Uses subadditive ergodic theory (Kingman's theorem) to prove the existence of a deterministic almost-sure growth exponent.
- Derives the statistical properties of burst sizes based on the residence-time distribution of the Markov chain.
Micro-Macro Correspondence: Proves that a population of regime-modulated Hawkes processes (microscopic self-exciting point processes) converges to the macroscopic Volterra limit, both in probability (annealed) and almost surely (quenched).
Numerical Experiments: Simulates the system on various network topologies (ring, star, Erdős-Rényi) to validate theoretical predictions regarding burst localization, spectral routing, and the annealed-quenched dichotomy.

4. Key Contributions and Results

A. The Annealed–Quenched Dichotomy

The paper establishes a sharp separation between two stability concepts:

Annealed Stability: The system is stable in expectation if the averaged memory gain $\bar{\rho} = \sum \pi_z \rho_z$ (where $\rho_z$ is the fractional branching ratio of regime $z$ ) is below a critical threshold. This yields uniform moment bounds.
Quenched Instability: Even if $\bar{\rho}$ is subcritical, typical sample paths can exhibit intermittency. Rare, long sojourns in a supercritical regime ( $\rho_z > 1$ ) are amplified by the long memory kernel, leading to macroscopic bursts.
Result: The system can be stable in expectation ( $\sup_t E\|U(t)\| < \infty$ ) while having a positive almost-sure growth exponent ( $\limsup_{t\to\infty} t^{-1}\log \|U(t)\| > 0$ ).

B. Heavy-Tailed Burst Statistics

The paper derives the statistical nature of the bursts:

Mechanism: The burst size $B$ is determined by the residence time $\tau_U$ in the unstable regime. If $\tau_U \sim \text{Exp}(q_{US})$ , the burst size follows a power-law tail:
$P(B > b) \sim b^{-q_{US}/\gamma_U}$
where $\gamma_U$ is the growth rate in the unstable regime.
Implication: This explains how systems can generate extreme events (heavy tails) even when average behavior appears benign.

C. Micro-Macro Limit

The authors prove that the burst mechanism is intrinsic to the macroscopic limit. A large population of microscopic Hawkes processes, modulated by the same regime switching, converges to the random-coefficient Volterra equation. The heavy-tailed burst statistics observed at the microscopic level persist in the macroscopic limit, confirming that these are not finite-population artifacts but structural features of the dynamics.

D. Geometric and Spectral Effects

Numerical experiments reveal that:

Burst Localization: Bursts are not uniform; they concentrate in specific spectral channels (Laplacian eigenmodes) and spatial nodes depending on network topology.
Non-Commutativity: Even when excitation operators do not commute (breaking diagonalizability), bursts remain concentrated in low-dimensional subspaces, though they exhibit mode mixing.

5. Significance and Impact

Theoretical Advancement: This work challenges the classical intuition that stability in switching systems can be inferred solely from averaged generators or frozen-regime criteria. It highlights that memory creates persistence, making the duration of regime visits a critical stability parameter.
Application Relevance: The findings are highly relevant for:
- Neuroscience: Modeling neuronal avalanches and spike-history effects.
- Finance: Explaining volatility clustering and heavy-tailed market crashes.
- Epidemiology: Understanding disease spread with history-dependent contacts.
- Climate Science: Analyzing regime shifts with long-term persistence.
Risk Assessment: The paper argues that relying solely on "annealed" (average) stability metrics is dangerous for systems with long memory. Risk assessments must incorporate "quenched" (pathwise) diagnostics to account for rare, persistent excursions that lead to catastrophic bursts.

6. Conclusion

The paper provides a rigorous mathematical framework for memory-driven intermittency. It demonstrates that the interplay between stochastic regime switching and long-range memory creates a unique instability mechanism where stability in expectation coexists with pathwise growth and heavy-tailed extreme events. This "annealed–quenched dichotomy" is specific to non-Markovian switching systems and necessitates new tools for stability analysis and risk management in complex networks.