Effective quenched linear response for random dynamical systems

This paper establishes effective quenched linear response for non-uniformly expanding random dynamical systems with non-i.i.d. compositions, enabling the proof of parameter differentiability for the CLT variance and annealed linear response through applications to sufficiently mixing base maps and various one- and higher-dimensional examples.

The Gist

Imagine you are the captain of a ship navigating through a stormy ocean. The ocean represents a random dynamical system—a world where the rules of physics change slightly every second due to random gusts of wind (the "randomness"). Your ship is the system itself, and the "weather" is the parameter you are tweaking (let's say, the angle of your sails).

For a long time, mathematicians could predict where your ship would end up on average if the weather was perfectly random (like flipping a coin every second). But what if the weather is "sticky"? What if a storm today makes a storm tomorrow more likely? This is a non-uniformly expanding system. It's chaotic, but the chaos isn't perfectly random; it has memory and structure.

This paper, written by D. Dragičević and Y. Hafouta, is like a new, high-tech navigation manual for these tricky, "sticky" storms. Here is the breakdown of their discovery in simple terms:

1. The Problem: The "Foggy" Forecast

In the past, if you wanted to know how your ship's path would change if you tilted your sails just a tiny bit, mathematicians could give you a general answer: "It will change smoothly." This is called Linear Response.

However, in these "sticky" storm scenarios, the old math had a major flaw. It could tell you the ship would change, but it couldn't tell you how fast or how reliably it would change. It was like a weather forecaster saying, "It might rain tomorrow," but refusing to say if it would be a drizzle or a hurricane, or if the rain would happen in 10 minutes or 10 days. The predictions were too vague to be useful for precise engineering or physics.

2. The Breakthrough: The "Effective" Compass

The authors developed a new method to get an "Effective" Linear Response.

Think of the old method as a compass that works, but the needle is wobbly and sometimes spins wildly before settling. The new method is a laser-guided compass. It doesn't just say "the ship moves"; it calculates the exact speed and direction of that movement with a high degree of precision.

They achieved this by proving that even in these messy, non-random storms, the system settles down fast enough that we can predict the outcome of small changes with mathematical certainty, not just vague probability.

3. Two Ways to Look at the Storm

The paper distinguishes between two ways of looking at your journey:

• The "Quenched" View (The Solo Sailor): This is looking at one specific storm path. "If this specific sequence of wind gusts happens, where will I be?" The authors proved that for almost every single specific storm path, you can predict the change precisely.
• The "Annealed" View (The Fleet Captain): This is looking at the average of all possible storms. "If I send out 1,000 ships into random storms, what is the average path?"
  • The Surprise: In the past, knowing the "Solo Sailor" path didn't guarantee you could predict the "Fleet Captain" average. The authors proved that with their new "laser-guided" math, if you can predict the solo path, you can also predict the fleet average. This is a huge leap forward.

4. The "Variance" Puzzle (The Bumpy Ride)

One of the most practical applications they found involves variance. In statistics, variance is a measure of how "bumpy" the ride is. If you are rolling dice, the variance tells you how much the numbers jump around.

In physics, knowing how the "bumpiness" of a system changes when you tweak a parameter is crucial.

• The Old Way: We knew the ride was bumpy, but we didn't know if making the sail angle slightly different would make the ride smoother or bumpier.
• The New Way: The authors proved that you can calculate exactly how the bumpiness changes when you tweak the system. They showed that the "bumpiness" is a smooth, predictable curve, not a jagged, unpredictable mess.

5. The "Magic" Ingredient: Mixing

How did they do it? They relied on a concept called Mixing.
Imagine stirring a cup of coffee with milk.

• Bad Mixing: The milk stays in a blob. You can't predict where a drop of milk will go.
• Good Mixing: The milk swirls and spreads out quickly. Even if you start with a specific drop, it quickly becomes part of the whole.

The authors focused on systems where the "milk" (the randomness) mixes well enough, even if it's not perfect. They showed that as long as the mixing happens at a certain speed (even if it's a bit slow at first), their "laser-guided" math works.

The Big Picture

In everyday language, this paper is a rulebook for predicting the unpredictable.

Before this, if you tried to build a bridge in a place with weird, unpredictable wind patterns, you might have to over-engineer it massively because you couldn't be sure how the wind would react to small changes in the bridge's design.

Now, thanks to this paper, we have a tool that says: "Even in these chaotic, non-random environments, if you make a small change, you can calculate exactly how the system will react, how the average behavior will shift, and how the 'bumpiness' of the system will change."

It turns a chaotic, foggy ocean into a map with clear, measurable currents, allowing scientists and engineers to navigate complex systems with much greater confidence.

Technical Summary

1. Problem Statement

The paper addresses the problem of linear response in the context of random dynamical systems (RDS). Specifically, it investigates the differentiability of statistical quantities (such as physical measures and variances) with respect to a perturbation parameter $\varepsilon$ in systems that are non-uniformly expanding.

• Context: In deterministic dynamics, linear response theory studies how the invariant measure $\mu_\varepsilon$ changes as the map $T_\varepsilon$ is perturbed. In random dynamics, the system is defined by a cocycle $(T_{\omega, \varepsilon})_{\omega \in \Omega}$ over a base probability space $(\Omega, \mathcal{F}, P, \sigma)$.
• The Challenge: Previous results on quenched linear response (differentiability for almost every realization $\omega$) for non-uniformly expanding systems (where expansion is not uniform across $\omega$) relied on the Multiplicative Ergodic Theorem (MET). This approach typically yields bounds involving tempered random variables (variables growing sub-exponentially).
• Limitation of Prior Work: Tempered bounds are insufficient for certain applications, such as proving the differentiability of the variance in the Central Limit Theorem (CLT) or establishing annealed linear response (differentiability of the averaged measure). The paper aims to overcome this by obtaining "effective" bounds where the controlling random variables belong to specific $L^p$ spaces ($p>0$), rather than just being tempered.

2. Methodology

The authors develop a new abstract framework that avoids the discretization of the parameter $\varepsilon$ and the direct reliance on the MET for each individual cocycle.

A. Abstract Framework (Banach Space Setup)

The core of the methodology is a set of abstract conditions formulated on a triplet of Banach spaces: $B_w \subset B_s \subset B_{ss}$ (Weak, Strong, and Stronger norms).

• Transfer Operators: The system is analyzed via transfer operators $L_{\omega, \varepsilon}$ acting on these spaces.
• Key Conditions: The authors establish conditions for:
  1. Uniform Boundedness: $\|L^n_{\sigma^{-n}\omega, \varepsilon}\|_w \leq C_0(\omega)$.
  2. Polynomial Decay of Correlations: $\|L^n_{\omega, \varepsilon} h\|_w \leq C_1(\omega) n^{-\beta} \|h\|_s$ for $h$ with zero mean.
  3. Differentiability of Operators: $\|(L_{\omega, \varepsilon} - L_\omega)h\|_w \leq C_2(\omega)|\varepsilon|\|h\|_s$.
  4. Second-Order Differentiability: Existence of a linear operator $\hat{L}_\omega$ approximating the derivative of $L_{\omega, \varepsilon}$.
  5. Existence of Equivariant Densities: A measurable family of densities $h_{\omega, \varepsilon}$ satisfying the invariance equation.
  6. Cone Contraction/Mixing: Conditions on the base dynamics ensuring sufficient mixing (e.g., $\psi$-mixing or $\alpha$-mixing).

B. "Effective" Bounds via Mixing Assumptions

Instead of using the MET to derive exponential decay with tempered constants, the authors utilize cone contraction arguments combined with specific mixing assumptions on the base process (the sequence of random variables driving the system).

• They prove that if the base process satisfies appropriate mixing conditions (e.g., polynomial decay of mixing coefficients) and moment conditions, the random variables controlling the decay rates (like $C_1, C_2$) belong to $L^p$ spaces for sufficiently large $p$.
• This allows them to replace the "tempered" error terms in previous works with $L^p$-integrable error terms.

C. Avoiding Discretization

Unlike previous approaches (e.g., Dragičević et al., 2019) which required discretizing $\varepsilon$ into a sequence $\varepsilon_k \to 0$ to handle the dependence of the invariant measure on $\varepsilon$, this paper handles the continuous parameter $\varepsilon$ directly. This is achieved by constructing the invariant densities as uniform limits of iterated transfer operators, ensuring the full-measure set of validity is independent of $\varepsilon$.

3. Key Contributions

1. Effective Quenched Linear Response (Theorem A):

   • Proves that for a class of non-uniformly expanding RDS, the map $\varepsilon \mapsto h_{\omega, \varepsilon}$ is differentiable in the weak norm.
   • Crucial Innovation: The error term in the linear response expansion is controlled by a random variable $U_1 \in L^s(\Omega)$ (for some $s>0$), rather than just a tempered variable. This "effectiveness" is the paper's main theoretical breakthrough.

2. Annealed Linear Response (Theorem B):

   • Establishes the differentiability of the annealed measure (the average over $\Omega \times M$).
   • This result was previously out of reach for non-uniformly expanding systems because the failure of annealed response was demonstrated in cases where only tempered bounds were available. The $L^p$ control allows the interchange of limits and integrals required for the annealed case.

3. Differentiability of the Variance in the Quenched CLT (Theorem C):

   • Applies the effective linear response to prove that the asymptotic variance $\Sigma^2_\varepsilon$ in the quenched Central Limit Theorem is differentiable with respect to $\varepsilon$.
   • This is the first result of its kind for systems with non-uniform decay of correlations. Previous results were limited to systems with uniform decay.

4. Concrete Applications (Theorem 13):

   • Verifies all abstract conditions for a wide class of one-dimensional random expanding maps (and certain higher-dimensional examples).
   • Demonstrates that if the base map has sufficient mixing and the random maps have sufficient regularity (e.g., $C^5$ dependence on parameters), the $L^p$ conditions are satisfied.

4. Main Results

• Theorem A (Abstract Quenched LR): Under specific integrability conditions on the random variables governing the system's expansion and mixing, the quenched linear response holds with an error bound in $L^s$.
  $$\left\| \frac{h_{\omega, \varepsilon} - h_{\omega}}{\varepsilon} - \hat{h}_\omega \right\|_w \leq U_1(\omega) |\varepsilon|^a$$
  where $U_1 \in L^s$ and $a > 0$.

• Theorem B (Annealed LR): The map $\varepsilon \mapsto \int \Phi d\mu_\varepsilon$ is differentiable at $\varepsilon=0$, where $\mu_\varepsilon$ is the annealed measure.

• Theorem C (Variance Differentiability): The variance $\Sigma^2_\varepsilon$ of the quenched CLT is differentiable at $\varepsilon=0$. The derivative is given by differentiating the sum of correlation terms in the variance formula.

• Theorem 13 (Application): For a class of random $C^5$ expanding maps on the circle (or interval) driven by a mixing base process, all conditions for Theorems A, B, and C are satisfied provided the base process has sufficient moment bounds and mixing rates.

5. Significance

• Bridging the Gap: The paper bridges the gap between the theory of linear response for uniformly hyperbolic systems (where results are robust) and non-uniformly hyperbolic systems (where results were previously limited to statistical stability or required tempered bounds).
• New Tools for Variance: By establishing $L^p$ bounds instead of tempered bounds, the authors unlock the ability to differentiate the variance in the CLT. This is critical for understanding the sensitivity of fluctuation statistics in random environments.
• Annealed Response: It resolves the question of whether annealed linear response holds for non-uniformly expanding systems, showing that it does, provided the base mixing is strong enough.
• Methodological Shift: The move away from the Multiplicative Ergodic Theorem (which forces discretization and tempered bounds) toward cone contraction and mixing-based estimates represents a significant methodological advancement in the analysis of random dynamical systems.

In summary, this paper provides a rigorous, "effective" framework for linear response in random systems with non-uniform expansion, enabling new applications to the differentiability of statistical limits like the variance in the CLT.