Gaussian concentration, integral probability metrics, and coupling functionals for infinite lattice systems

This paper establishes a transport-entropy framework for Gaussian concentration on infinite lattice systems, proving that while the associated transportation costs lack metric extensivity, the integral probability metric and coupling functional coincide in finite volumes to extend the Kantorovich-Rubinstein theorem and characterize concentration via Marton's inequality, ultimately converging to the $\bar d$-metric in the thermodynamic limit.

The Gist

The Big Picture: Measuring Chaos in a Giant Grid

Imagine a massive, infinite checkerboard stretching out in every direction. On every single square of this board, there is a tiny switch that can be either ON or OFF. This is your "infinite lattice system."

Now, imagine you have a rulebook (a probability measure) that tells you how likely it is for the switches to be ON or OFF. Sometimes, the switches are totally random and independent (like flipping a coin for every square). Other times, they are connected: if one switch is ON, its neighbors are more likely to be ON too (like a crowd of people all wearing the same color shirt).

The Goal: The authors want to understand how much the "average behavior" of this giant grid can fluctuate. If you look at a huge section of the grid, how likely is it that the average is very different from what you expect?

In math, this is called Concentration. If the system is "concentrated," it means the average is very stable and rarely wanders far from the expected value.

The Problem: The "Ruler" Doesn't Work

For small, finite grids (like a 10x10 checkerboard), mathematicians have a standard tool to measure these fluctuations. It's like using a ruler (a metric) to measure the distance between two different patterns of switches.

• The Old Way (Bobkov-Götze Theorem): If you want to know how "far apart" two patterns are, you count how many switches are different. This works great for small grids. It's like saying, "Pattern A is 5 switches different from Pattern B."

• The New Problem (Infinite Grids): The authors discovered that when you try to apply this "ruler" to an infinite grid, it breaks.

  • Analogy: Imagine trying to measure the "distance" between two infinite oceans by counting how many drops of water are different. The number is infinite! The ruler becomes useless.
  • The Discovery: The specific type of stability (Gaussian concentration) that happens in these infinite systems cannot be measured by any standard ruler or cost function. The "distance" between patterns in this context is a weird, non-standard shape that doesn't fit into the usual boxes of geometry.

The Solution: A New Kind of "Transport"

Since the old ruler doesn't work, the authors invented two new tools to measure the relationship between different patterns. Think of these as two different ways to describe the same journey.

1. The "Integral Probability Metric" (The Judge)

Imagine a Judge who looks at two different crowds (two probability distributions). The Judge asks: "What is the biggest difference in opinion between these two crowds on any topic?"

• If the crowds are very similar, the Judge finds no big differences.
• If they are very different, the Judge finds a huge gap.
• This tool measures the gap by looking at how the crowds react to various "questions" (functions).

2. The "Coupling Functional" (The Matchmaker)

Imagine a Matchmaker trying to pair up people from Crowd A with people from Crowd B.

• The Matchmaker wants to pair them up so that as many people as possible are identical (both ON or both OFF).
• The "cost" is how many pairs end up being different (one ON, one OFF).
• The Matchmaker tries to find the best possible pairing to minimize the number of mismatches.

The Big Breakthrough (The Duality):
The authors proved a magical theorem: The Judge's score and the Matchmaker's cost are exactly the same.

• No matter how you look at it (asking questions or pairing people), the "distance" between the two crowds is identical.
• This is a modern, super-charged version of a famous old math theorem (Kantorovich-Rubinstein), but it works even when the "ruler" doesn't exist.

The "Thermodynamic Limit": Zooming Out to Infinity

The paper also looks at what happens when you zoom out from a small square of the grid to the entire infinite universe.

• The Scaling Trick: If you just count the total differences in an infinite grid, the number is infinite. So, the authors divide the difference by the size of the area they are looking at. It's like measuring "density" instead of "total weight."
• The Result: When they do this, all their weird new tools (the Judge and the Matchmaker) converge to a single, famous concept in physics called the $\bar{d}$-distance.
• What is $\bar{d}$? It's a standard way in physics to measure how different two infinite patterns are, essentially asking: "If I look at a random spot, how likely is it that the two patterns disagree?"

The Final Connection: Entropy and Stability

The paper concludes by linking this new "distance" to Entropy (a measure of disorder or surprise).

• The Rule: If two patterns are "close" (low $\bar{d}$-distance), they must have low "relative entropy" (they are very similar in their randomness).
• The Implication: This proves that if a system has a certain type of stability (Gaussian concentration), it forces the system to be unique. You can't have two different stable states that look the same on average. This helps physicists prove that certain materials don't have "phase transitions" (like water turning to ice) under specific conditions.

Summary in a Nutshell

1. The Issue: Standard math tools (rulers) fail to measure stability in infinite, connected grids.
2. The Fix: The authors created two new, non-standard tools (a "Judge" and a "Matchmaker") to measure the distance between patterns.
3. The Discovery: These two tools are actually the same thing (Duality).
4. The Result: As you zoom out to infinity, these tools settle into a known physics concept ($\bar{d}$-distance), proving that stability in these systems is deeply connected to how much "disorder" (entropy) exists.

Creative Metaphor:
Think of the infinite grid as a giant, chaotic dance floor.

• Old Math: Tried to measure the chaos by counting every single step difference. (Impossible, infinite steps).
• New Math: Instead of counting steps, they asked: "If we try to pair up dancers from two different groups, how many pairs are dancing to different beats?"
• The Finding: They proved that the "beat difference" (entropy) perfectly predicts how chaotic the dance floor is, even without being able to count the steps.

Technical Summary

1. Problem Statement and Motivation

The paper addresses the study of Gaussian concentration inequalities for random fields on infinite product spaces $\Omega = S^{\mathbb{Z}^d}$, where $S$ is a finite set (e.g., spin systems in statistical mechanics).

• The Context: In finite-dimensional systems, Gaussian concentration is often linked to transportation-cost inequalities via the Bobkov–Götze theorem. This theorem characterizes concentration for Lipschitz functions using the $T_1$ inequality, which relates the Wasserstein distance (induced by a metric) to the relative entropy.
• The Obstacle: In infinite-dimensional lattice systems, the natural sensitivity measure for local functions is the $\ell_2$-norm of local oscillations ($\|\delta f\|_2$), rather than a standard Lipschitz constant with respect to a metric on the configuration space.
• The Core Question: Can the Gaussian concentration bound involving $\|\delta f\|_2$ be characterized by a transportation inequality induced by a metric on the configuration space?
• The Finding: The authors prove that no such metric exists. The transportation structure naturally associated with $\ell_2$-Gaussian concentration is non-extensive and cannot be generated by any metric or cost function on $S^{\mathbb{Z}^d}$. This necessitates a new framework beyond classical optimal transport.

2. Methodology and Definitions

The authors develop a transport-entropy framework specifically tailored for infinite product spaces, avoiding reliance on a global metric.

Key Definitions

1. Uniform $\ell_2$-Gaussian Concentration Bound (GCB):
   A measure $\mu$ satisfies GCB with constant $C$ if for all local functions $f$:
   $$\log \int e^{f - \mu(f)} d\mu \leq \frac{C}{2} \|\delta f\|_2^2$$
   where $\delta_i f$ is the oscillation of $f$ at site $i$, and $\|\delta f\|_2 = (\sum_i (\delta_i f)^2)^{1/2}$.

2. Integral Probability Metric (IPM) - $D_{p,\Lambda}$:
   For a finite volume $\Lambda$, defined via the dual formulation:
   $$D_{p,\Lambda}(\nu, \mu) = \sup_{f \in \text{Loc}(\Omega_\Lambda) \setminus \text{Const}} \frac{\nu(f) - \mu(f)}{\|\delta f\|_q}$$
   where $p$ and $q$ are conjugate exponents ($1/p + 1/q = 1$). This measures the difference in expectations normalized by the oscillation norm.

3. Generalized Kantorovich Functional (GKF) - $Q_{p,\Lambda}$:
   A coupling-based functional defined as:
   $$Q_{p,\Lambda}(\mu, \nu) = \left( \inf_{\Pi \in \mathcal{C}_\Lambda(\mu, \nu)} \sum_{i \in \Lambda} \Pi(\sigma_i^{(1)} \neq \sigma_i^{(2)})^p \right)^{1/p}$$
   Unlike the standard Wasserstein distance (which sums the cost of mismatches), this sums the probabilities of mismatches raised to the power $p$.

3. Key Contributions and Results

A. Finite Volume Duality (The Main Structural Result)

The paper's central theorem (Theorem 4.1) establishes an exact duality in finite volumes:
$$D_{p,\Lambda}(\nu, \mu) = Q_{p,\Lambda}(\mu, \nu)$$

• Significance: This extends the classical Kantorovich–Rubinstein duality. While the classical theorem equates IPMs and Wasserstein distances induced by a metric, this result equates an IPM (based on oscillations) with a coupling functional that is not induced by any metric on the configuration space.
• Characterization of Concentration: Consequently, the uniform $\ell_2$-Gaussian concentration bound is equivalent to the inequality:
  $$Q_{2,\Lambda}(\mu, \nu) \leq \sqrt{2C s_\Lambda(\nu|\mu)}$$
  where $s_\Lambda$ is the relative entropy. This provides a new characterization of concentration for dependent random variables via Marton's coupling inequality.

B. Non-Metricity and Non-Extensivity (Appendix A)

The authors rigorously prove why the classical Bobkov–Götze framework fails in this setting:

1. Non-Metricity: The $\ell_2$-norm of oscillations cannot be controlled by any Lipschitz constant associated with a cost function on the configuration space.
2. Non-Extensivity: If one attempts to force a metric structure (e.g., using an $\ell_2$-type distance on configurations), the resulting concentration bounds fail to be extensive. The constants required to bound the concentration blow up as the volume $\Lambda \to \mathbb{Z}^d$, rendering the inequality trivial or useless in the thermodynamic limit.

C. Thermodynamic Limits and the $\bar{d}$-Metric

The paper analyzes the behavior of these functionals in the infinite-volume limit for translation-invariant measures:

• Rescaling: The authors show that the limits of the normalized metrics exist:
  $$D_p(\nu, \mu) = \lim_{n \to \infty} \frac{D_{p,\Lambda_n}(\nu, \mu)}{|\Lambda_n|^{1/p}}$$
• Coincidence with $\bar{d}$: Remarkably, for all $p \geq 1$, these limits coincide with the $\bar{d}$-metric (Ornstein's distance) from ergodic theory:
  $$D_p(\nu, \mu) = \bar{d}(\nu, \mu) = \inf_{\Pi} \Pi(\sigma_0^{(1)} \neq \sigma_0^{(2)})$$
  where the infimum is over jointly translation-invariant couplings.
• Implication: While the finite-volume functionals differ from standard Wasserstein distances, their thermodynamic limits converge to the same fundamental metric of ergodic theory.

D. Thermodynamic Gaussian Concentration Bounds

The authors introduce a Thermodynamic Gaussian Concentration Bound (T-GCB) involving the relative entropy density $s(\nu|\mu)$ and the $\bar{d}$-distance:
$$\bar{d}(\mu, \nu) \leq \sqrt{2C s(\nu|\mu)}$$

• Equivalence: For asymptotically decoupled measures (a class including Gibbs measures in the uniqueness regime), the T-GCB is equivalent to the transport-entropy inequality above.
• Weak Version: They also define a "weak" version based on limsup of finite-volume pressures, which holds for any translation-invariant measure and implies the transport-entropy inequality with the lower relative entropy density.

4. Significance and Impact

1. Beyond Metric Frameworks: The paper demonstrates that transport-entropy structures in infinite-dimensional statistical mechanics are richer than those captured by metric-based optimal transport. It identifies a specific "transport structure" that is intrinsic to the $\ell_2$-oscillation sensitivity but not generated by a metric.
2. Unification of Concepts: It bridges the gap between:
   • Concentration Inequalities (McDiarmid, Marton).
   • Transport Inequalities (Wasserstein, Kantorovich).
   • Ergodic Theory ($\bar{d}$-metric).
3. New Characterization: It provides a rigorous equivalence between Gaussian concentration and coupling inequalities for random fields, extending Marton's results to the infinite-volume setting without requiring a metric on the configuration space.
4. Phase Transitions: The results offer tools to study phase transitions. Specifically, the transport-entropy inequality involving the $\bar{d}$-distance implies the uniqueness of the translation-invariant Gibbs measure (absence of phase transition) for asymptotically decoupled systems.

Summary Conclusion

Chazottes, Collet, and Redig successfully construct a transport-entropy theory for infinite lattice systems that respects the natural $\ell_2$-sensitivity of local functions. By proving that this structure cannot be metric-induced, they introduce generalized Kantorovich functionals that coincide with integral probability metrics in finite volume and converge to the $\bar{d}$-metric in the thermodynamic limit. This work fundamentally extends the scope of concentration inequalities and transport inequalities beyond the classical metric setting.