Maximum Cluster Diameter in Non-Critical Bond Percolation

This paper establishes that in non-critical Bernoulli bond percolation for dimensions $d \ge 2$, the maximum diameter of finite clusters scales asymptotically as $\varkappa(p) \log n$ almost surely, where the constant $\varkappa(p)$ is determined by the exponential decay rate of large cluster probabilities, and further analyzes the asymptotic behavior of the number of vertices in such large-diameter clusters.

The Gist

Imagine a giant, infinite grid made of city blocks (like a 3D chessboard). In this city, every street connecting two blocks has a chance of being open or closed. If a street is open, you can walk across it; if it's closed, you can't. This is the world of bond percolation.

The paper by Kaito Kobayashi asks a very specific question about this city: How big can the biggest "island" of connected blocks get if we are not at the exact tipping point where the whole city suddenly connects?

Here is the breakdown of the paper's findings using simple analogies:

1. The Setting: The "Just Right" vs. The "Off"

In this model, there is a special "tipping point" probability (called $p_c$).

• At the tipping point: The city is chaotic. You might have a massive island that stretches forever, or tiny islands everywhere. It's a critical, messy state.
• Away from the tipping point (The focus of this paper): The author looks at two scenarios:
  • Too few open streets: The islands are small and isolated.
  • Too many open streets: There is one giant, infinite island that covers the whole city, but there are also many small, isolated "islands" floating in the gaps.

The paper ignores the giant infinite island and focuses entirely on the largest of the small, finite islands within a square box of size $n$.

2. The Main Discovery: The "Logarithmic" Growth Rule

The author measures the "diameter" of these islands (how far you have to walk from one end to the other).

The Finding:
If you keep making your city box bigger and bigger (increasing $n$), the size of the biggest finite island doesn't grow linearly (like $n$). Instead, it grows very slowly, following a logarithmic curve.

The Analogy:
Imagine you are looking for the tallest tree in a forest that keeps getting bigger.

• If you double the size of the forest, the tallest tree doesn't double in height.
• The paper proves that the tallest tree grows at a predictable, steady pace relative to the logarithm of the forest size.
• Specifically, the size of the biggest island is roughly $\kappa \times \log(n)$.
  • $n$ is the size of the box.
  • $\log(n)$ is the "slow growth" factor.
  • $\kappa$ is a constant number that depends on how likely the streets are to be open.

The paper calculates exactly what this constant $\kappa$ is. It is determined by how quickly the probability of finding a connection drops off as you get further away. Think of it as the "decay rate" of connectivity.

3. The "What If" Scenarios (Large Deviations)

The paper also asks: What are the odds that we find an island that is much bigger than the usual "logarithmic" size?

The Finding:
If you look for an island that is, say, twice as big as the typical maximum, the probability of finding it is extremely low.

• The paper provides a formula to calculate exactly how rare these "giant outliers" are.
• Analogy: If the typical tallest tree in a forest of 1 million trees is 50 feet, finding a 100-foot tree is possible but incredibly rare. The paper gives you the exact mathematical odds of finding that 100-foot tree.

4. Counting the "Big" Islands

Finally, the paper looks at how many people (or vertices) live on these unusually large islands.

The Finding:
Even though these large islands are rare, the paper shows that the number of people living on them follows a very predictable pattern.

• Analogy: If you count how many people live in the "top 1%" of the largest islands in your city, the paper proves that this count is very stable. If you repeat the experiment many times, the number of people you count will almost always be very close to the average prediction.

Summary of the "Takeaway"

In a world where connections are random but not at the chaotic tipping point:

1. Size Limit: The biggest isolated group of connected items grows very slowly (logarithmically) as the space gets larger.
2. Predictability: We can calculate the exact speed of this growth based on how "sticky" the connections are.
3. Rarity: Finding a group significantly larger than this limit is exponentially rare.
4. Stability: The number of items in these rare, large groups is highly predictable and consistent.

The paper essentially draws a precise map of the "geography" of these random islands, telling us exactly how big the biggest ones can get and how often we might see a giant outlier.

Technical Summary

Technical Summary: Maximum Cluster Diameter in Non-Critical Bond Percolation

Problem Statement
This paper investigates the geometric characteristics of finite connected components in independent (Bernoulli) bond percolation on the $d$-dimensional integer lattice $\mathbb{Z}^d$ ($d \geq 2$) within the non-critical regime ($p \neq p_c$). While the existence and uniqueness of infinite clusters, critical values, and correlation decay are well-established, the sharp asymptotics of geometric quantities for finite clusters in non-critical regimes remain less understood. Specifically, the paper addresses the behavior of the maximum diameter, denoted $R_n$, of finite clusters contained within a box $B_n$ of side length $2n+1$. The diameter is measured using the $\ell_\infty$ metric. The study focuses on determining the almost sure asymptotic growth rate of $R_n$ as $n \to \infty$, establishing large deviation principles for deviations from this typical size, and analyzing the spatial extent (number of vertices) of clusters exhibiting large diameters.

Methodology
The analysis relies on the exponential decay of connectivity probabilities in the non-critical regime. The author utilizes the decay rate $\xi(p)$, defined as:
$$\xi(p) = \lim_{n \to \infty} \left\{ -\frac{1}{n} \log P_p(0 \leftrightarrow \partial B_n, |C_0| < \infty) \right\}$$
where $\xi(p) = \phi(p)$ for $p < p_c$ and $\xi(p) = \sigma(p)$ for $p > p_c$. The constant governing the diameter scaling is defined as $\kappa(p) = d/\xi(p)$.

The proofs adapt techniques from van der Hofstad and Redig regarding the maximum volume of finite clusters, extending them to the diameter. Key methodological steps include:

1. Borel-Cantelli and Boole's Inequalities: Used to establish almost sure convergence and bound probabilities of unions of events.
2. Discretization and Independence: To prove lower bounds and large deviation principles, the box $B_n$ is partitioned into a grid of smaller, well-separated sub-boxes ($A_n$). This ensures that the events of clusters having specific diameters in these sub-boxes are independent, allowing for product estimates.
3. Variance Analysis: For the asymptotic behavior of the number of vertices in large-diameter clusters, the author employs Chebyshev's inequality. This involves a detailed estimation of the variance of the count variable $S_n(\rho)$, splitting the covariance terms based on the distance between vertices to exploit the exponential decay of correlations.
4. Boundary Conditions: The paper distinguishes between free boundary conditions ($R_n^{(fb)}$) and zero boundary conditions ($R_n^{(zb)}$, where bonds outside $B_n$ are forced closed), proving their asymptotic equivalence.

Key Results
The paper establishes four main theorems:

• Theorem 2.1 (Almost Sure Convergence): For $p \neq p_c$, the maximum diameter $R_n$ satisfies:
  $$\frac{R_n}{\log n} \to \kappa(p) \quad \text{almost surely as } n \to \infty.$$
  This confirms that the maximum diameter grows logarithmically with the box size, with the rate determined by the inverse of the correlation length decay rate scaled by dimension.

• Theorem 2.2 (Boundary Condition Equivalence): The difference between the maximum diameter under free boundary conditions and zero boundary conditions vanishes asymptotically:
  $$\lim_{n \to \infty} P_p(R_n^{(zb)} \neq R_n^{(fb)}) = 0.$$

• Theorem 2.3 (Large Deviation Principle): For any $\rho > \kappa(p)$, the probability that the maximum diameter exceeds $\rho \log n$ decays exponentially in $\log n$. Specifically, the rate function $\varsigma(\rho)$ exists and is given by:
  $$\varsigma(\rho) = \lim_{n \to \infty} -\frac{1}{\log n} \log P_p(R_n > \rho \log n) = \xi(p)\rho - d.$$
  This quantifies the rarity of observing clusters significantly larger than the typical scale.

• Theorem 2.4 (Asymptotics of Vertex Count): Let $S_n(\rho)$ be the number of vertices in $B_n$ belonging to finite clusters with diameter exceeding $\rho \log n$ (where $0 < \rho < \kappa(p)$). The paper proves that $S_n(\rho)$ concentrates around its mean:
  $$\frac{S_n(\rho)}{E_p[S_n(\rho)]} \xrightarrow{P_p} 1.$$
  Furthermore, the expected number of such vertices scales as:
  $$n^{d - \xi(p)\rho - \varepsilon} \leq E_p[S_n(\rho)] \leq n^{d - \xi(p)\rho + \varepsilon}.$$

Significance and Claims
The paper claims to fill a gap in the understanding of finite cluster geometry in non-critical percolation. While the tail behavior of cluster sizes and radii has been studied extensively, sharp asymptotics for the maximum diameter were previously unknown. The results provide a precise characterization of the spatial extent of the largest finite clusters, linking this geometric quantity directly to the exponential decay rate of connectivity.

The author notes that the framework extends previous work on maximum cluster volume (van der Hofstad and Redig) to diameter. Additionally, the paper suggests that these results may be extended to diameters measured in chemical (graph) distance, citing recent work by Dembin and Nakajima on rate functions for chemical distance, though this extension is not proven within the current text. The work remains strictly theoretical, focusing on rigorous probabilistic bounds and asymptotic limits without proposing experimental applications or future physical models.