Quenched Amplification and Tail Shaping in Networked Systems with Memory and Regime Switching

1. Problem Statement

The paper addresses a critical phenomenon in large-scale networked dynamical systems (e.g., cyber-physical infrastructures, distributed sensing) operating under intermittent adverse conditions and long-term memory.

• The Paradox: These systems often appear stable in aggregate (annealed stability) based on mean-field diagnostics or second-moment estimates. However, they exhibit rare but extreme trajectory-level excursions (bursts) that are invisible to standard stability criteria.
• The Gap: Existing theories for switching systems with memory often focus on "annealed" properties (expectations over all regime paths). They fail to control "quenched" behavior (the distribution of maxima for a fixed realization of the switching path).
• The Mechanism: The paper investigates how the combination of stochastic regime switching (Markovian changes in operating conditions), long memory (modeled via Volterra-type convolution), and non-normal network geometry (directional amplification) generates heavy-tailed transient bursts, even when individual regimes are stable.

2. Methodology

A. Mathematical Framework: Lifted Switched ODE

To handle the non-Markovian nature of memory, the authors employ a Sum-of-Exponentials (SOE) approximation of the completely monotone memory kernel. This allows the infinite-dimensional Volterra integral equation to be embedded into a finite-dimensional lifted Ordinary Differential Equation (ODE).

• State Vector: $X(t) = (x(t), y_1(t), \dots, y_K(t))^\top$, where $x$ is the network state and $y_k$ are auxiliary memory states.
• Dynamics: $\dot{X}(t) = A^{(m(t))}_{z(t)} X(t) + \tilde{f}(t)$, where $z(t)$ is a continuous-time Markov chain (regime) and $m(t)$ is a control mode.
• Key Insight: Even if the instantaneous network operator $B_z$ is stable, the lifted operator $A_z$ can exhibit transient growth due to the interaction between memory accumulation and non-normal geometry.

B. Analytical Approach: Annealed vs. Quenched

The paper establishes a sharp dichotomy:

1. Annealed Boundedness: On finite horizons, the system is bounded in expectation ($\sup E[\|X(t)\|] < \infty$) due to standard Grönwall arguments and bounded forcing.
2. Quenched Heavy Tails: For fixed regime paths, rare long dwell times in an "unfavorable" regime ($U$) allow memory to accumulate coherently. If the system aligns with an amplifying direction (cone condition), this leads to exponential growth during the dwell, resulting in power-law tails for the burst size distribution.

C. Control Strategy: DDDAS

The authors propose a Dynamic Data-Driven Applications Systems (DDDAS) intervention strategy.

• Indicators: Two real-time indicators are monitored:
  1. Latent Memory Load ($L(t)$): Accumulated energy in memory states.
  2. Spectral Susceptibility ($S(t)$): The Euclidean logarithmic norm (matrix measure) of the lifted operator, indicating instantaneous growth potential.
• Policy: A hysteresis-based controller switches the system to "Verify" or "Mitigate" modes when indicators exceed thresholds.
• Action: In "Mitigate" mode, the system switches to a strictly contractive operator (increasing damping, accelerating memory decay) for a minimum dwell time, effectively truncating the amplification pathway without altering the exogenous regime statistics.

3. Key Contributions

1. Identification of a Structural Mechanism: The paper proves that heavy tails in memory-driven switching systems arise from a specific competition: the persistence of unfavorable regimes (exponential dwell times) vs. the effective lifted growth rate ($\gamma_U$) induced by memory geometry.
2. Rigorous Tail Bounds:
   • Derives a lower bound for the burst tail probability: $P(B_T > b) \ge C b^{-\lambda_U / \gamma_U}$.
   • The tail exponent is explicitly determined by the ratio of the unfavorable regime exit rate ($\lambda_U$) to the effective growth rate ($\gamma_U$).
3. Computable Early-Warning Indicators:
   • Introduces $S(t)$ (logarithmic norm of the lifted operator) as a measurable proxy for instability, distinct from eigenvalue analysis.
   • Demonstrates that $S(t) > 0$ signals the activation of a dangerous growth channel even when the nominal system is stable.
4. Tail Shaping via Contraction-on-Demand:
   • Proves that the DDDAS policy can truncate the tail (making the burst size bounded) or steepen the tail exponent (reducing risk) by enforcing contraction only when indicators signal danger.
   • Crucially, this is achieved without modifying the statistics of the external regime switching process.

4. Results

Numerical Validation

The authors simulate a directed, non-normal network with periodic forcing and SOE memory.

• Annealed vs. Quenched Separation: Simulations show that while the mean and median trajectories remain bounded and periodic, the upper quantiles (e.g., 99th percentile) drift upward, confirming the existence of heavy tails.
• Indicator Performance: The indicators $L(t)$ and $S(t)$ rise systematically prior to amplification events, validating their use as early warnings.
• Tail Truncation:
  • Uncontrolled: The Complementary Cumulative Distribution Function (CCDF) of burst sizes follows a power law with slope $\approx -\lambda_U/\gamma_U$.
  • DDDAS Controlled: The CCDF shifts downward and steepens. In strong mitigation scenarios, the tail is visibly truncated (finite maximum burst), matching the theoretical prediction of pathwise boundedness.
• Baseline Comparison: The DDDAS approach outperforms "memory-free" baselines (which remove the diagnostic channel) and "SAFE-in-U" baselines (which permanently alter regime dynamics), as it preserves typical system behavior while selectively suppressing extremes.

5. Significance and Implications

• Theoretical Advancement: The work bridges the gap between heavy-tail theory (Kesten-Goldie results) and non-modal stability analysis. It demonstrates that heavy tails can emerge in linear, deterministic systems driven by bounded forcing, provided there is memory and regime switching, challenging the notion that heavy tails require multiplicative noise or nonlinearity.
• Operational Relevance:
  • Risk Management: Provides a framework for understanding why systems can be "stable on average" yet prone to catastrophic failures.
  • Control Design: Offers a practical, data-driven control strategy (DDDAS) that mitigates extreme events without over-damping the system or altering the underlying stochastic environment.
  • Diagnostics: The use of the logarithmic norm of the lifted operator provides a concrete, computable metric for monitoring "hidden" instability in memory-rich systems.

In summary, the paper provides a geometrically grounded explanation for extreme events in memory-driven networked systems and offers a provably effective, adaptive control mechanism to shape and mitigate tail risk.