Finitary coding and Gaussian concentration for random fields

This paper establishes that Gaussian concentration inequalities are preserved under finitary codings of i.i.d. fields provided the coding volume has a finite second moment (or first moment under specific structural assumptions), thereby deriving sharp necessary and sufficient conditions for such concentration in classical lattice models like Ising and Potts systems and characterizing geometric ergodicity in one-dimensional processes.

The Gist

Imagine you are trying to predict the weather in a massive, complex city. You have a super-powerful computer (the i.i.d. field) that generates random, independent weather reports for every single street corner. These reports are pure chaos: the rain on 5th Avenue has absolutely nothing to do with the sun on 6th Avenue. They are completely independent.

Now, imagine you want to create a realistic weather map for the city where things do make sense. If it's raining heavily on 5th Avenue, it's likely raining on 6th Avenue too. You need a rulebook (a coding) to translate those random, independent reports into a coherent, dependent city-wide weather system.

This paper is about understanding the rules of that translation process and how "predictable" the final city weather will be.

The Core Concept: The "Finitary" Rulebook

In math, we call this translation a finitary coding. Here is the catch:

• The Old Way: To know the weather at one spot, you might have to look at the entire history of the universe. That's impossible.
• The Finitary Way: To know the weather at a specific spot, you only need to look at a finite (limited) neighborhood of the random reports.

However, there's a twist: How big is that neighborhood?
Sometimes, to figure out the weather at a specific house, you only need to look at the 3 houses next door. Other times, a weird storm front might require you to look at the whole block, or even the whole city. The size of the "look-ahead" zone is random. We call this the Coding Radius.

The Big Question: How Stable is the System?

The authors are asking: If we build a complex system (like a weather map or a magnetic material) using these finite look-aheads, how stable is it?

They are looking for Gaussian Concentration. In plain English, this means: If you make a small change to the input, does the output stay roughly the same?

• Good Concentration: If you change the weather report on one street, the overall city forecast doesn't go crazy. The fluctuations are small and predictable (like a bell curve).
• Bad Concentration: A tiny change in one spot causes a massive, unpredictable ripple effect across the whole city.

The Main Discovery: The "Size of the Look-Ahead" Matters

The paper proves a beautiful relationship between the size of the neighborhood you look at and the stability of the system.

Think of the "Coding Volume" as the total number of street corners you have to check to figure out the weather for one spot.

1. The "Safe" Zone (Finite Second Moment):
   If the average size of your look-ahead neighborhood is small enough (specifically, if the square of the size is finite), then the system is stable. It will have good Gaussian concentration. Small changes in input lead to small, predictable changes in output.

   • Analogy: Imagine you are building a house of cards. If the "look-ahead" is small, you are only stacking a few cards high. It's stable.

2. The "Special" Zone (Finite First Moment):
   The authors found that if the system has a specific structure (like a "short-range factorization," which happens in certain computer algorithms called Coupling-from-the-Past), you don't need the neighborhood to be that small. You just need the average size to be finite.

   • Analogy: If your house of cards is built with a special locking mechanism (the structural assumption), you can stack it a bit higher (larger neighborhoods) and it will still stand.

3. The "Danger" Zone (Infinite Moments):
   If the look-ahead neighborhoods get too huge on average (infinite first or second moment), the system becomes unstable. You lose Gaussian concentration.

   • Analogy: If you have to look at the entire city to decide the weather for one house, a tiny change in a distant city could flip a switch and cause a hurricane in your backyard. The system is too sensitive.

Real-World Examples: Why Should We Care?

The authors apply this to famous problems in physics and math:

• The Ising Model (Magnets): Imagine a grid of tiny magnets. They can point Up or Down.

  • High Temperature (Uniqueness Regime): The magnets are jiggling around. There is only one way the system behaves. The "look-ahead" is small. Result: The system is stable and predictable.
  • Low Temperature (Phase Coexistence): The magnets want to align. You can have a state where everything is Up, or everything is Down. Here, the "look-ahead" becomes infinite. Result: The system is unstable. A tiny nudge can flip the whole magnet grid.
  • The Critical Point: This is the exact moment of transition. The paper shows that at this exact point, the "look-ahead" is so huge (infinite average size) that the system loses its stability, even though a rulebook technically exists.

• Parking Cars: Imagine cars trying to park on a street one by one. If the street is jammed, you have to look far back to see if a spot is open. The paper shows that if the "jamming" isn't too extreme, the final parking pattern is stable and predictable.

The "Aha!" Moment

The most exciting part of the paper is that it unifies many different areas.

• Before, scientists had different tools for magnets, parking problems, and computer algorithms.
• Now, they have one master key: The size of the look-ahead neighborhood.

If you can prove that the "look-ahead" isn't too big (mathematically, that its average size or squared average size is finite), you automatically know the system is stable and predictable.

Summary in a Metaphor

Imagine a Game of Telephone played by a million people.

• The Input: A whisper starts at one person.
• The Coding: Each person whispers what they heard to their neighbors.
• Finitary: Each person only listens to a limited number of neighbors before speaking.

The Paper's Conclusion:
If the number of neighbors each person listens to is generally small (finite moments), then the final message at the other end of the line will be a clear, predictable version of the start.
But, if the "listening radius" grows uncontrollably (infinite moments), the final message will be a chaotic mess, and a tiny change in the starting whisper will completely destroy the final sentence.

The authors have given us the mathematical ruler to measure that "listening radius" and tell us exactly when the game of telephone will work and when it will fail.

Technical Summary

1. Problem Statement

The paper investigates the relationship between Gaussian concentration inequalities and the structural properties of finitary codings in random fields.

• Gaussian Concentration: A property where local observables of a random field exhibit sub-Gaussian fluctuations with constants independent of the size of their dependence sets. This is a strong form of stability and mixing.
• Finitary Codings: A mechanism where a dependent random field $Y$ is generated as a shift-equivariant image of an i.i.d. (independent and identically distributed) field $X$, such that the value of $Y$ at any site depends on a finite but random portion of $X$.
• The Core Question: Under what conditions is Gaussian concentration preserved when an i.i.d. field is transformed into a dependent field via a finitary coding? Specifically, what moment conditions on the coding volume (the size of the random region of the input required to determine an output) are necessary and sufficient?

2. Methodology

The authors employ a combination of probabilistic inequalities, ergodic theory, and structural analysis of random fields.

• Refinement of Bounded-Differences Inequalities: The proof relies on a refinement of the Talagrand-Marton inequality. Unlike classical McDiarmid's inequality, which assumes deterministic Lipschitz constants, this version accommodates configuration-dependent influences. The influence of an input variable on the output is treated as a random variable dependent on the coding radius.
• Truncation and Telescoping: To handle the random coding radius, the authors introduce a truncated coding map $\phi^{(n)}$ with a deterministic radius $n$. They analyze the difference $|f(\phi^{(n)}(x)) - f(\phi^{(n)}(x'))|$ by telescoping over output coordinates and bounding the resulting indicators using the overlap of random coding windows.
• Convolution Analysis: The key technical step involves bounding the sum of squared influence coefficients. This is reduced to analyzing the $\ell_1$ norm of a convolution term involving the coding radii. The authors use Young's inequality for discrete convolutions to relate this to the moments of the coding volume.
• Short-Range Factorization: For a weaker condition (finite first moment), the authors introduce a structural assumption called the short-range factorization property. This property, satisfied by Coupling-From-The-Past (CFTP) constructions, allows the correlation between coding windows to be factorized, reducing the required moment condition.
• Sharpness Analysis: The authors construct counterexamples (e.g., the critical Ising model) to demonstrate that their moment conditions are optimal and cannot be relaxed without additional structural assumptions.

3. Key Contributions and Results

A. Abstract Theoretical Results

1. Second-Moment Condition (Theorem 3.1):

   • If a random field $Y$ is a finitary coding of an i.i.d. field $X$, and the coding volume has a finite second moment ($E[|B_\infty(0, r_\phi)|^2] < \infty$), then $Y$ satisfies Gaussian concentration.
   • This is the general condition required for arbitrary finitary codings.

2. First-Moment Condition under Structure (Theorem 3.3):

   • If the coding satisfies the short-range factorization property (satisfied by CFTP constructions), then a finite first moment ($E[|B_\infty(0, r_\phi)|] < \infty$) is sufficient for Gaussian concentration.
   • This significantly broadens the applicability to models where the second moment might diverge but the first moment is finite.

3. Sharpness of Conditions:

   • The paper proves that these conditions are sharp. If the expected coding volume is infinite, Gaussian concentration can fail.
   • Example: The critical Ising model ($d \ge 2$) admits a finitary coding from an i.i.d. field, but the expected coding volume is infinite, and Gaussian concentration fails.

4. Structural Consequences (Theorem 2.1 & 2.2):

   • Bernoullicity: For finite-valued shift-invariant fields, Gaussian concentration implies the field is isomorphic to a Bernoulli shift (a strong mixing property).
   • Positive Relative Entropy: Gaussian concentration implies that any distinct ergodic measure has positive relative entropy with respect to the field's law. This is used to prove the failure of concentration in phase-coexistence regimes.

B. Applications to Specific Models

The authors apply these abstract results to unify and extend previous knowledge in statistical mechanics and stochastic processes:

• Gibbs Measures on $\mathbb{Z}^d$ (Ising, Potts, Random-Cluster):

  • Result: Gaussian concentration holds if and only if the model is in the full uniqueness regime (no phase coexistence).
  • Significance: Previous results were limited to strict sub-regimes (e.g., Dobrushin uniqueness or high-temperature limits). This paper covers the entire uniqueness regime, including models previously inaccessible to older methods.
  • Criticality: At the critical point, concentration fails because the susceptibility diverges, leading to infinite expected coding volume.

• Probabilistic Cellular Automata (PCA):

  • Uniformly ergodic PCAs satisfy the short-range factorization property. If the coding volume has a finite first moment, they satisfy Gaussian concentration.

• One-Dimensional Processes:

  • Markov Chains: For countable-state, irreducible, aperiodic chains, Gaussian concentration is equivalent to:
    1. Geometric ergodicity.
    2. Exponential return-time tails.
    3. Existence of a finitary i.i.d. coding with exponential tails.
  • Chains with Unbounded Memory: The results extend to $g$-measures. If a CFTP construction exists with finite expected regeneration time, the chain satisfies Gaussian concentration.

4. Significance and Impact

• Unified Framework: The paper provides a single theoretical framework that links the structural generation of a random field (finitary coding) with its statistical stability (Gaussian concentration).
• Optimality: It establishes precise boundaries for when concentration holds, showing that the "uniqueness regime" in statistical mechanics is exactly the regime where these fields can be finitarily coded with controlled volume.
• Beyond Classical Regimes: By utilizing recent constructions of finitary codings (e.g., by Spinka et al.), the authors extend concentration results to models that were previously out of reach for techniques like Dobrushin's uniqueness criterion or disagreement percolation.
• New Characterizations: In one dimension, the work offers a complete characterization of geometric ergodicity in terms of finitary coding and concentration, bridging ergodic theory and concentration of measure.
• Open Problems: The paper identifies open questions, such as whether Gaussian concentration implies the existence of a finitary coding (the converse) and the behavior of models with polynomially decaying coding tails.

Conclusion

This paper demonstrates that Gaussian concentration is a robust property preserved under finitary codings, provided the "cost" of the coding (the volume of the input required) has finite moments. The work fundamentally links the geometric properties of coding radii to the probabilistic behavior of the resulting fields, offering sharp necessary and sufficient conditions for a wide class of statistical mechanical and stochastic models.