Asymptotics of the overlap distribution of branching Brownian motion at high temperature

This paper investigates the precise decay rate of the probability of non-zero overlap between two particles in high-temperature branching Brownian motion, revealing that the transition between two sub-phases occurs at different inverse temperature thresholds depending on whether the analysis is conditioned on the branching process or performed unconditionally.

The Gist

Imagine a vast, invisible forest growing in a computer simulation. This isn't a normal forest; it's a Branching Brownian Motion (BBM).

Here's how it works:

1. The Seed: At the start, one single particle (a "seed") appears at the center of a line.
2. The Walk: It wanders randomly left and right, like a drunk person walking home.
3. The Split: After a random amount of time, it splits into two identical children.
4. The Repeat: Each child starts wandering from where the parent stopped, and eventually, they too split.
5. The Explosion: Over time, this creates a massive, branching tree of particles, with millions of "leaves" (particles) at the end.

Now, imagine we are physicists looking at this forest. We want to understand the Gibbs Measure. Think of this as a "popularity contest."

• Particles that wander far to the right (high energy) are considered "winners."
• We assign them a higher probability of being picked. If you pick a particle at random based on this popularity contest, you are much more likely to pick a "winner" than a "loser."

The Big Question: How Similar Are Two Random Winners?

The paper asks: If I pick two "winners" from this forest at random, how much of their history did they share?

• The Overlap: Imagine two hikers starting at the same tree. They walk together for a while, then split off. The "overlap" is the percentage of the total journey they spent walking side-by-side before splitting.
• The Temperature: The "inverse temperature" ($\beta$) is a knob we can turn.
  • High Temperature (Low $\beta$): The contest is loose. Even particles that didn't wander very far right have a decent chance of being picked. The system is "chaotic."
  • Low Temperature (High $\beta$): The contest is fierce. Only the absolute furthest-right particles get picked. The system is "ordered."

The paper focuses on the High Temperature phase. Intuitively, you might think: "If the contest is loose, the winners are all over the place, so two random winners probably split off very early and have almost zero overlap."

The paper confirms this: Yes, as time goes on, the overlap drops to zero. But the authors wanted to know: How fast does it drop? And more importantly, does the answer change depending on how we look at it?

The Two Ways of Looking (The Twist)

The paper reveals a surprising twist. There are two ways to calculate this probability, and they give different answers!

1. The "Typical" View (The Real Forest)

This asks: "If I look at one specific forest that actually grew, what is the chance two winners share a path?"

• The Analogy: You are standing in a specific forest. You pick two winners.
• The Result: The probability of them sharing a path decays at a certain speed.
• The Surprise: There is a "tipping point" at a specific temperature setting ($\beta = \sqrt{2}/2$).
  • Below the tipping point: The winners are like a huge crowd of people who all split off early. They behave like a normal crowd.
  • Above the tipping point: The winners are "extremists." They are the rare particles that managed to stay near the very top of the tree. Their behavior changes drastically, and the math describing them shifts to a different, more complex pattern (involving "stable distributions," which are like heavy-tailed storms).

2. The "Average" View (The Statistical Average)

This asks: "If I simulate a million forests and average the results, what is the chance?"

• The Analogy: You are a statistician looking at a spreadsheet of a million forests.
• The Result: The probability decays at a different speed, and the tipping point is in a different place ($\beta = \sqrt{2}/3$).
• Why the difference? This is the most fascinating part.
  • In the "Typical" view, you see what usually happens.
  • In the "Average" view, you are being tricked by rare, weird forests.
  • Imagine a forest where, by pure luck, one single particle wandered incredibly far to the right. In that specific forest, everyone picked as a "winner" is related to that one lucky particle. They all share a huge overlap!
  • These "lucky forests" are incredibly rare, but when they happen, the overlap is huge. When you average them into the "Mean," these rare giants pull the average up, making it look like the overlap is higher than it usually is.
  • The paper shows that the "Average" view is dominated by these rare, lucky events, while the "Typical" view ignores them.

The "Drift" Metaphor

To visualize why the overlap changes, imagine the particles are boats on a river.

• The River: Represents time.
• The Current: Represents the "drift" (the tendency to move right).
• The Split: When a boat splits, the two new boats go their own way.

In the "Typical" view:

• At low temperatures, the boats just drift with the current. They split early, and the overlap is zero.
• At the first tipping point, the boats that matter are the ones that managed to swim against the current to stay near the top. They have to be very careful not to drift too far down. This changes their behavior.

In the "Average" view:

• The math is trickier. The "Average" is heavily influenced by the rare scenario where a boat gets a massive, lucky boost (a "lucky current") that pushes it way ahead of everyone else.
• Because of this, the "Average" view sees a different tipping point where the strategy for being a "winner" changes from "swim normally" to "ride the lucky current."

Why Does This Matter?

This isn't just about math trees. This model is used to understand:

1. Spin Glasses: Materials where atoms are confused and can't agree on a direction (like a messy crowd).
2. Polymers: Long chains of molecules folding in a solution.
3. High-Energy Physics: How particles scatter when they smash into each other.

The paper teaches us a profound lesson about Statistics vs. Reality:

• Reality (Typical): Usually, things are boring and follow the standard rules.
• Statistics (Average): The "average" can be completely misleading because it is dragged around by rare, extreme outliers.

In the world of branching particles, the "average" forest is not the same as the "typical" forest. The paper maps out exactly where these two worlds diverge, showing us that the "temperature" at which the system changes its behavior depends entirely on whether you are looking at a single forest or the average of a million.

Technical Summary

1. Problem Statement and Context

The paper investigates the overlap distribution of particles in a Branching Brownian Motion (BBM) in the high-temperature phase (subcritical regime, $\beta < \sqrt{2}$).

• The Model: In a BBM, particles perform Brownian motions and split into two at rate 1. The positions of particles at time $t$ are interpreted as energy levels.
• The Gibbs Measure: Particles are selected according to a Gibbs measure $G_{\beta, t}$ with inverse temperature $\beta$, where the weight of a particle $u$ is proportional to $e^{\beta X_u(t)}$.
• The Overlap: For two particles $u, v$ chosen independently from $G_{\beta, t}$, the overlap $q_t(u, v)$ is the fraction of time they shared a common ancestor (rescaled by $t$).
• The Question: While it is known that the overlap converges to 0 almost surely for $\beta < \sqrt{2}$, the paper seeks the precise decay rate of the probability that the overlap exceeds a threshold $a \in (0, 1)$. Specifically, it analyzes:
  1. The typical overlap distribution ($\nu_{\beta, t}$): The conditional probability given the realization of the BBM.
  2. The mean overlap distribution ($\mathbb{E}[\nu_{\beta, t}]$): The unconditional probability.

The authors aim to determine the exponential decay rates and polynomial corrections for these probabilities as $t \to \infty$.

2. Methodology

The authors employ a combination of probabilistic techniques tailored to branching processes and Gaussian fields:

• Spinal Decomposition and Change of Measure: A central tool is the change of measure (introduced by Chauvin and Rouault) that transforms the BBM into a process with a distinguished "spine" particle.
  • Under the measure $Q_\beta$, the spine performs a Brownian motion with drift $\beta$ and splits at rate 2.
  • The paper utilizes a specific change of measure $Q_{2\beta}$ to analyze the numerator of the overlap expression, which involves terms like $e^{2\beta X_u(t)}$.
• Additive Martingales: The analysis relies heavily on the convergence properties of the additive martingale $W_t(\beta) = e^{-\psi(\beta)t} \sum e^{\beta X_u(t)}$. The paper distinguishes between regimes where $W_t(\beta)$ is uniformly integrable and where it is not.
• Extremal Process: For higher temperatures (closer to criticality), the behavior is dominated by particles near the maximum of the BBM. The authors use the convergence of the extremal process to a Decorated Poisson Point Process (DPPP).
• Brownian Estimates: Precise estimates for Brownian motions conditioned to stay below barriers (ballot theorems) are used to derive the polynomial corrections ($t^{-1/2}, t^{-3/2}$).
• Stable Laws: In the lower temperature regime, the limits involve $\alpha$-stable random variables arising from the supercritical additive martingale behavior.

3. Key Contributions and Results

The paper reveals that the decay behavior splits into distinct sub-phases, and surprisingly, the thresholds differ between the typical and mean overlap distributions.

A. Typical Overlap Distribution (Theorem 1.1)

The decay rate of $\nu_{\beta, t}([a, 1])$ depends on $\beta$:

1. Upper Temperature ($0 \le \beta < \sqrt{2}/2$):
   • The decay is exponential: $e^{-(1-\beta^2)at}$.
   • The limit is a constant times the ratio of martingale limits $W_\infty(2\beta)/W_\infty(\beta)^2$.
   • Heuristic: The Gibbs measure at $2\beta$ is subcritical; the contributing particles behave like a Brownian motion with drift $2\beta$ up to time $at$.
2. Critical Temperature ($\beta = \sqrt{2}/2$):
   • The decay is $e^{-at/2}$ with a $\sqrt{t}$ correction.
   • The limit involves the derivative martingale $Z_\infty$.
3. Lower Temperature ($\sqrt{2}/2 < \beta < \sqrt{2}$):
   • The decay is $e^{-(\sqrt{2}-\beta)^2 at}$.
   • The limit involves a non-degenerate $(\sqrt{2}/2\beta)$-stable random variable $S_{2\beta}$.
   • Heuristic: The contribution comes from the extremal process (particles near the top), leading to stable fluctuations.

Threshold for Typical Overlap: The transition occurs at $\beta = \sqrt{2}/2$.

B. Mean Overlap Distribution (Theorem 1.2)

The decay rate of $\mathbb{E}[\nu_{\beta, t}([a, 1])]$ exhibits a different phase structure:

1. Infinite Temperature ($\beta = 0$):
   • Decay: $2at e^{-at}$.
2. Upper Temperature ($0 < \beta < \sqrt{2}/3$):
   • Decay: $e^{-(1-\beta^2)at}$.
   • The limit is the expectation of the typical limit (Remark 1.3).
3. Critical Temperature ($\beta = \sqrt{2}/3$):
   • Decay: $t^{-1/2} e^{-at/3}$.
4. Lower Temperature ($\sqrt{2}/3 < \beta < \sqrt{2}$):
   • Decay: $t^{-3/2} e^{-(2-\beta^2)^2 at / 8\beta^2}$.
   • Heuristic: The mean is dominated by a rare event where a particle achieves an atypically large position. The optimal drift for this particle changes from $2\beta$ to $1/\beta + \beta/2$ at the threshold.

Threshold for Mean Overlap: The transition occurs at $\beta = \sqrt{2}/3$.

4. Significance and Physical Interpretation

• Discrepancy in Thresholds: The most striking result is that the phase transition for the mean overlap ($\beta = \sqrt{2}/3$) is different from the typical overlap ($\beta = \sqrt{2}/2$).
  • For $\beta \in [\sqrt{2}/3, \sqrt{2})$, the mean overlap is governed by rare fluctuations (large deviations) where a single particle contributes disproportionately to the numerator of the overlap expression.
  • For $\beta < \sqrt{2}/3$, the mean is determined by the typical behavior.
• Connection to Spin Glasses: The results provide rigorous mathematical validation for predictions in spin glass theory (specifically the Generalized Random Energy Model - GREM). The overlap distribution is crucial for understanding the ultrametric structure of the Gibbs measure.
• High-Energy Physics: The paper connects these findings to diffractive scattering in high-energy particle physics, where the overlap distribution relates to the statistics of rapidity gaps.
• Methodological Advance: The paper successfully handles the delicate issue of the denominator $W_t(\beta)^2$ in the mean overlap calculation, showing that for $\beta > 0$, the "small denominator" events do not dominate the mean, unlike the $\beta=0$ case.

5. Conclusion

Chataignier and Pain provide a complete asymptotic description of the overlap distribution for BBM in the high-temperature phase. They identify two distinct regimes for the typical overlap and three for the mean overlap, separated by critical inverse temperatures $\sqrt{2}/2$ and $\sqrt{2}/3$ respectively. The work highlights the complex interplay between typical behavior and large deviations in disordered systems, offering precise exponential rates and polynomial corrections that refine previous conjectures in statistical physics.