Level 2.5 large deviations and uncertainty relations for non-Markov self-interacting dynamics

This paper establishes a general framework for the exact level 2.5 large deviation statistics of non-Markovian self-interacting jump processes by exploiting timescale separation, thereby enabling the derivation of generalized thermodynamic and kinetic uncertainty relations for trajectory observables in such systems.

The Gist

Imagine you are watching a crowd of people walking through a giant, empty train station.

In a normal (Markovian) world, how a person walks depends only on where they are right now. If they are at the ticket counter, they might walk to the platform. If they are at the platform, they might walk to the exit. Their past doesn't matter; they have no memory.

But in the real world (and in this paper), things are different. Imagine these people are "self-interacting." As they walk, they leave sticky notes on the floor.

• If a person has spent a lot of time near the ticket counter, they leave a trail of notes there.
• Future people (or even the same person later) see these notes and think, "Oh, everyone hangs out here, I should go there too!" or "This place is crowded, let's avoid it."

This is a Self-Interacting Jump Process (SIJP). The system's future behavior depends on its entire history. It's like a conversation where you keep reminding everyone of what was said ten minutes ago, changing the tone of the conversation as it goes on.

The Big Problem: Predicting the Rare

Scientists love to study "average" behavior. But what about the rare, weird events?

• What if, by pure chance, everyone suddenly decides to run to the exit at the exact same time?
• What if the crowd gets stuck in a loop forever?

In normal systems, we have a mathematical "rulebook" (Large Deviation Theory) to predict how likely these rare events are. But for these "memory-having" systems, the rulebook was missing. We didn't know how to calculate the odds of these strange, history-dependent fluctuations.

The Paper's Solution: The "Level 2.5" Map

The authors (Francesco, Amarjit, and Juan) have written a new rulebook. They call it "Level 2.5."

Think of it like a weather forecast:

• Level 1: "It will rain tomorrow." (Too simple).
• Level 2: "It will rain 50% of the time." (Better, but still vague).
• Level 2.5: "Here is the exact map of where the rain will fall and how fast the wind will blow, considering how the clouds from yesterday are pushing today's storm."

They figured out how to calculate the joint probability of two things happening simultaneously over a long time:

1. Where the system spent its time (The "Empirical Measure"): Did the crowd hang out at the ticket counter or the exit?
2. How often they jumped between places (The "Empirical Flux"): How many people ran from A to B?

The Secret Sauce: The "Slow-Motion" Trick
How did they solve this? They noticed a clever trick in time.

• The "sticky notes" (the memory) change very slowly. It takes a long time for the crowd to build up a big trail.
• The people walking (the jumps) move very fast.

Because the memory changes slowly, for a split second, the system acts like a normal, memory-less system. The authors used this "separation of timescales" to freeze the memory, solve the problem for that split second, and then stitch all those moments together to get the full picture.

The Result: Uncertainty Rules

Once they had this new map, they derived two powerful "laws of physics" for these memory-based systems. These are called Uncertainty Relations.

Think of these as a trade-off between Precision and Cost.

1. The Kinetic Uncertainty Relation (The "Energy Bill"):

   • The Metaphor: Imagine you want to measure the speed of a car very precisely.
   • The Rule: To get a super-precise measurement, the car must be moving frantically (high activity). If the car is lazy and moving slowly, your measurement will be fuzzy.
   • The Paper's Twist: Even if the car is "remembering" its past speed, this rule still holds. If you want a precise reading of a system's activity, the system must be busy. You can't have a quiet, precise system.

2. The Thermodynamic Uncertainty Relation (The "Heat Bill"):

   • The Metaphor: Imagine trying to keep a cup of coffee at a perfect temperature.
   • The Rule: To keep the temperature steady (low fluctuation), you have to waste a lot of energy (entropy) heating and cooling it constantly.
   • The Paper's Twist: Even if the coffee cup has a "memory" of how hot it was yesterday, the rule stands: Precision costs energy. If you want to reduce the randomness of a flow (like traffic or electricity), you have to pay a "tax" in the form of wasted heat or entropy.

Why Does This Matter?

This isn't just about math. This applies to:

• Ants: How they leave pheromone trails to find food.
• Bacteria: How they navigate chemical gradients.
• AI and Robots: How we build artificial agents that learn from their past to make better decisions.

The paper tells us that no matter how smart or "memory-rich" these systems are, they are still bound by fundamental laws. If you want them to be precise and reliable, they have to work hard and burn energy. You can't cheat the universe, even with a good memory.

In short: The authors built a new mathematical telescope that lets us see exactly how systems with "memories" behave when they do something rare. They proved that even with memory, nature still demands a price for precision.

Technical Summary

1. Problem Statement

The paper addresses a fundamental gap in statistical physics: the lack of a closed-form formulation for dynamical large deviations (LDs) in non-Markovian systems.

• Context: While dynamical LDs and their corollaries (such as fluctuation bounds and uncertainty relations) are well-established for Markovian systems, general results for non-Markovian dynamics are scarce, often limited to semi-Markov processes.
• Specific Focus: The authors focus on Self-Interacting Jump Processes (SIJPs). These are continuous-time jump processes where the transition rates depend on the system's history via a functional of an empirical observable (e.g., the time-averaged occupation of states). This creates long-time memory and strong correlations, leading to phenomena like ergodicity breaking and modified first-passage times.
• Goal: To derive the exact "Level 2.5" large deviation rate function for SIJPs (joint statistics of empirical measure and flux) and use this to derive generalized Thermodynamic (TUR) and Kinetic (KUR) Uncertainty Relations for non-Markovian systems.

2. Methodology

The authors employ a combination of path integral techniques, time-scale separation arguments, and variational calculus.

• Model Definition:
  • A SIJP is defined on a discrete state space $S$ with jump rates $Q_{xy}(A_t)$ that depend on a state-dependent empirical observable $A_t = t^{-1} \int_0^t f(X_{t'}) dt'$.
  • The dependence of rates on the trajectory history ($A_t$) renders the process non-Markovian.
• Key Technical Advance: Time-Scale Separation:
  • The authors observe a separation of timescales at long times ($T \to \infty$).
  • The "microscopic" jump dynamics relaxes to a quasi-stationary state much faster than the empirical observable $A_t$ changes significantly.
  • This allows the system to be treated as a sequence of instantaneous Markov processes where the generator $Q$ evolves slowly based on the accumulated history.
• Derivation of Level 2.5 LDs:
  • The path probability is formulated with time-dependent rates and non-exponential waiting times.
  • The path measure is exponentially "tilted" (similar to the Markov case) to study rare fluctuations.
  • A change of variables ($t \to e^t$) and time-reversal ($t \to T-t$) are applied to handle the memory kernel.
  • The result is a variational problem minimizing a functional over auxiliary time-dependent distributions $\rho(t)$ and generators $H(t)$.
• Derivation of Uncertainty Relations:
  • Using the variational nature of the Level 2.5 rate function, the authors apply specific ansatzes (assuming the empirical density coincides with the stationary state) to derive upper bounds on the rate functions for specific observables (fluxes and currents).
  • These bounds are then expanded to second order to obtain inequalities relating variance, mean, and entropy production/activity.

3. Key Contributions and Results

A. Exact Level 2.5 Large Deviation Rate Function

The central result is the derivation of the joint large deviation rate function $I_{2.5}(\ell, \phi)$ for the empirical measure $\ell$ and empirical flux $\phi$.
$$I_{2.5}(\ell, \phi) = \inf_{(\rho, H)} I[\rho, H]$$
where the functional $I[\rho, H]$ is given by:
$$I[\rho, H] = \sum_{(x,y) \in E} \int_0^\infty e^{-t} \rho_x(t) Q_{xy}(M_t) \Gamma\left( \frac{H_{xy}(t)}{Q_{xy}(M_t)} \right) dt$$

• Variables: $\rho(t)$ is an auxiliary time-dependent distribution, $H(t)$ is an auxiliary time-dependent generator, and $M_t$ is the dynamical observable associated with $\rho(t)$.
• Significance: This generalizes the standard Level 2.5 LD for Markov jump processes. Unlike the Markov case, the contraction over $(\rho, H)$ can induce non-convexity, suggesting richer fluctuation behaviors in SIJPs. The factor $e^{-t}$ acts as a temporal discount, suppressing contributions from early times (which correspond to distant future times in the original frame).

B. Generalized Kinetic Uncertainty Relation (SIJP-KUR)

The authors derive a bound for the fluctuations of generic flux observables $B_T$:
$$\frac{\text{var}(b)}{b^2} \ge \frac{1}{k_{\text{SIJP}}}$$

• $k_{\text{SIJP}}$: The SIJP dynamical activity, defined as the average number of configuration changes per unit time, generalized to the non-Markovian case via an integral over the auxiliary trajectory.
• Implication: Precision in measuring a flux is bounded by the system's dynamical activity, even when memory effects are present.

C. Generalized Thermodynamic Uncertainty Relation (SIJP-TUR)

For current observables (antisymmetric fluxes), a refined bound is derived:
$$\frac{\text{var}(j)}{j^2} \ge 2 \int_0^\infty e^{-t} \frac{(j^\rho_t / j)^2}{\sigma_t} dt$$

• $\sigma_t$: The instantaneous entropy production rate.
• Interpretation: The integral represents a weighted average where the "cost" of precision is paid in terms of instantaneous entropy production. The $e^{-t}$ factor implies that entropy production at different times contributes differently to the bound, reflecting the system's memory horizon.

4. Illustrative Examples

The theory is validated using two specific models:

1. Two-State SIJP: A spin system where the flip-up rate grows exponentially with past occupation ($Q_{01} = 1 + e^{\alpha A_t}$).
   • Results show that the exact rate function $I_2(\ell)$ matches Monte Carlo simulations.
   • For large deviations, the non-Markovian character becomes apparent, deviating from the Markovian approximation due to sharp variations in the optimal trajectory $\rho(t)$.
2. Three-State Ring with Self-Induced Drag: A particle on a ring where transition rates slow down based on the time spent in a specific state.
   • Demonstrates the SIJP-TUR bound for particle currents, showing that the bound holds even with non-zero stationary currents and complex memory effects.

5. Significance and Future Directions

• Theoretical Unification: This work bridges the gap between Markovian large deviation theory and non-Markovian self-interacting systems, providing a rigorous framework for analyzing rare events in systems with memory.
• Universal Principles: It establishes that fundamental trade-offs (Uncertainty Relations) between precision, dissipation, and activity persist in non-Markovian regimes, albeit with modified, history-dependent forms.
• Applications: The framework is applicable to:
  • Biological Systems: Modeling autochemotactic organisms (e.g., E. coli, ants) that leave chemical trails.
  • Active Matter: Designing strategies for artificial self-propelling agents.
  • Quantum Systems: The authors suggest potential extensions to open quantum dynamics with memory.
• Limitations: While the results are exact in the statistical physics sense (large $T$ limit), they currently lack the full mathematical rigor of specific affine cases (Ref [31]), a gap the authors aim to close in future work.

In summary, the paper provides a powerful, general toolkit for quantifying fluctuations and thermodynamic costs in systems where the past dictates the future, extending the reach of modern non-equilibrium statistical mechanics beyond the Markovian approximation.