van den Berg-Kesten--type correlation inequalities for disjoint polymers in the KPZ universality class

This paper establishes a van den Berg-Kesten-type correlation inequality for the KPZ line ensemble and continuum directed random polymer by leveraging the integrability of the log gamma polymer and the geometric RSK correspondence, while demonstrating that such an inequality fails for non-integrable models.

The Gist

The Big Picture: A Game of "Disjoint Paths"

Imagine you are playing a game on a grid (like a giant chessboard). You have a bunch of hikers trying to walk from the bottom of the board to the top.

• The Environment: The board is covered in random "weather" (some spots are sunny and easy to walk, others are stormy and hard).
• The Goal: The hikers want to find the path with the best total weather (the "energy" or "weight" of the path).
• The Rule: The hikers cannot step on the same square. They must stay disjoint (separate) from each other.

This paper is about a specific mathematical rule called the BK Inequality. In simple terms, this rule asks: "If I know that one hiker found a really great path, does that make it more or less likely that a second, separate hiker will also find a great path?"

In the world of "zero temperature" (where hikers are super-efficient and only care about the single best path), the answer is known: They are negatively correlated. If the first hiker takes the "best" path, they use up all the good weather, leaving the second hiker with worse options. Knowing one did well makes it less likely the other did well.

The Problem: The "Positive Temperature" Twist

The authors are studying a more complex version of this game called Positive Temperature.

• The Metaphor: Imagine the hikers are now a bit "drunk" or "confused." Instead of picking just the single best path, they wander around a bit. They explore many different paths.
• The Consequence: The "score" isn't just the best path anymore; it's an average of all the paths they took, weighted by how good they were. This is called the Free Energy.

Here is the catch: In this "drunk" version, the old rule (the BK inequality) breaks.
Why? Because of Entropy (or "crowding").
In the zero-temperature game, if the first hiker takes a specific route, they block that route for the second. But in the positive-temperature game, the "score" depends on every possible path the hikers could have taken. Even if the first hiker's path looks great, the second hiker might still find a great score because they are exploring a huge "cloud" of possibilities, not just one line. The old logic of "blocking" doesn't work cleanly because the randomness is everywhere.

What the Authors Did

The authors, Ganguly, Hegde, and Zhang, wanted to prove a new version of this inequality for the "drunk" (positive temperature) hikers. They wanted to show that even in this messy, entropic world, there is still a way to say that two separate groups of hikers don't "help" each other too much.

The Challenge:
They couldn't just copy the old proof. The math for the "drunk" hikers is much harder because of that "entropy" factor. If they tried to force the old rule, it would fail.

The Solution: The "Log-Gamma" Trick
To solve this, they didn't work directly with the messy "drunk" hikers. Instead, they used a special, simpler version of the game called the Log-Gamma Polymer.

• The Analogy: Think of the Log-Gamma model as a "training simulator" for the real game. It's a discrete, step-by-step version of the problem where the math is "integrable" (meaning we have exact formulas for the answers, like having a cheat sheet).
• The Tool: They used a mathematical magic trick called the Geometric RSK correspondence. This is like a translator that converts the problem of "hikers on a grid" into a problem of "stacking blocks" or "line ensembles" (lines of numbers that interact with each other).

The Breakthrough:
Using this translator and the "cheat sheet" of the Log-Gamma model, they proved that:

1. If you condition on the first group of hikers (fix their path), the second group's performance is still "dominated" by a fresh, unconditioned group.
2. However, there is a catch. Because of the "entropy" (the crowd of possibilities), the second group's score needs to be shifted down by a small amount (a logarithmic shift) to make the inequality hold.
3. They also proved that if you try to use this rule for other types of random weather (distributions that aren't Log-Gamma), the rule fails. This highlights that the special "integrable" math of the Log-Gamma model was crucial to making the proof work.

The Main Results (Translated)

1. The Inequality: They proved that for the "drunk" hikers (the KPZ line ensemble), if you know the first hiker did very well, the second hiker is unlikely to do too well, provided you adjust for the "crowding" (entropy) by subtracting a small logarithmic amount from the second hiker's score.
2. The Error Margin: The rule isn't perfect; there is a tiny chance it fails (an error term), but that chance is so small it's practically zero (exponentially small).
3. The Application: They didn't just prove this for fun. They showed that this new inequality is the "missing key" needed to solve two other big problems in the field:
   • Calculating the probability of "upper tail" events (how likely is it for the hikers to find an incredibly good path?).
   • Proving that these hikers eventually look like "Brownian bridges" (a specific type of random curve) when conditioned on finding a great path.

Why This Matters (According to the Paper)

The paper emphasizes that this is a correction and a completion of previous work.

• Earlier papers tried to use a "naive" version of this rule for the "drunk" hikers, but the proof was flawed because it ignored the entropy issue.
• This paper fixes that flaw. It shows exactly how the rule works (with the shift) and proves it rigorously using the Log-Gamma model.
• It also serves as a warning: You can't just assume this rule works for any random system. It relies heavily on the special mathematical properties of the Log-Gamma model. If you change the rules of the game (the distribution of the weather), the inequality might break.

Summary Analogy

Imagine you are trying to predict the performance of two separate teams in a chaotic, noisy stadium.

• Old Rule (Zero Temp): If Team A finds the perfect seat, Team B definitely won't find a good one.
• New Rule (Positive Temp): Because the stadium is chaotic, Team A finding a good seat doesn't automatically ruin Team B's chances, but it does make it slightly less likely, if you account for the fact that Team B is juggling many more options (entropy).
• The Paper's Contribution: The authors built a special "simulation" (Log-Gamma) to prove exactly how much less likely Team B is to succeed, correcting previous attempts that got the math wrong. They showed that this specific simulation is the only way to get the proof to work.

Technical Summary

Technical Summary: Van den Berg–Kesten–Type Correlation Inequalities for Disjoint Polymers in the KPZ Universality Class

Problem Statement
The van den Berg–Kesten (BK) inequality is a fundamental tool in classical percolation theory, establishing that disjoint events are negatively correlated. This inequality has been successfully extended to zero-temperature models, specifically Last Passage Percolation (LPP) and the parabolic Airy line ensemble, where it serves as a crucial input for analyzing upper tail probabilities and geodesic scaling limits. However, extending this inequality to positive-temperature polymer models, such as the Continuum Directed Random Polymer (CDRP) and the associated KPZ line ensemble, presents significant conceptual obstacles.

In zero-temperature LPP, the energy of a path is determined solely by the random variables along its trajectory, allowing for "disjoint certificates" of events. In positive-temperature polymer models, the free energy (log-partition function) involves an integral over all paths, incorporating the entire environment's randomness. Consequently, the largeness of a second curve in the KPZ line ensemble cannot be tied to a simple disjoint certificate in the same way, as entropy plays a fundamental role. The authors note that a direct generalization of the zero-temperature BK inequality is unlikely to hold, necessitating a new formulation that accounts for entropic shifts.

Methodology
The authors prove a version of the BK inequality for the KPZ line ensemble and the CDRP by leveraging the integrability of the discrete log-gamma polymer model. The proof strategy proceeds in three main stages:

1. Discrete Log-Gamma Polymer Analysis: The authors work within the discrete log-gamma polymer model on $\mathbb{Z}^2$, an integrable model where exact formulas exist for joint distributions. They utilize the geometric Robinson-Schensted-Knuth (gRSK) correspondence to map the polymer partition functions to a discrete line ensemble.
2. Extended Invariance Identity: A critical technical component is the proof of an "extended invariance identity" (Proposition 2.10). While standard identities relate disjoint path partition functions to line ensembles when endpoints lie on the top boundary, the authors extend this to allow endpoints on the right boundary. This involves introducing an auxiliary line ensemble and a specific graph structure to handle paths crossing between different boundary conditions.
3. Stochastic Domination via Reweighting: Using the extended identity, the authors relate the partition function of two disjoint polymers to a single polymer in a smaller environment. They demonstrate that, conditional on the top line of the line ensemble, the law of the smaller partition function can be expressed as a reweighting of the original law. Crucially, this reweighting factor is shown to be negatively associated with increasing events via the FKG inequality. This yields a stochastic domination result for the discrete model (Theorem 2.1).
4. Scaling Limit: Finally, the authors pass to the continuum limit. They establish the weak convergence of the rescaled log-gamma partition functions to the CDRP partition functions ($Z$) and the multi-point partition functions ($K_2$). By carefully managing the entropy factors (Vandermonde determinants) and modulus of continuity estimates, they transfer the discrete inequality to the continuum setting.

Key Contributions and Results

• Theorem 2.1 (Discrete BK Inequality): The authors establish a BK-type inequality for the log-gamma polymer. Specifically, for disjoint endpoints, the conditional probability of the second curve (normalized by the first) falling into an increasing set is bounded by the unconditional probability of the first curve falling into that set. This result relies heavily on the integrability of the log-gamma model; the authors provide a counter-example (Remark 2.5) showing the inequality fails for non-integrable distributions like Bernoulli or Uniform weights, highlighting that integrability is essential for the validity of the result in the positive-temperature setting.
• Theorem 1.5 (Continuum Disjoint Polymer Inequality): This theorem provides a BK inequality for the continuum directed random polymer. It states that for disjoint starting and ending points, the log-partition function of two disjoint paths, conditioned on the first path's behavior, is stochastically dominated by a single path's partition function (up to a shift).
• Theorems 1.3 and 1.4 (KPZ Line Ensemble Inequality): These are the main results for the KPZ line ensemble ($\hat{h}_t$). They establish that conditioned on the first curve $\hat{h}_{t,1}$, the second curve $\hat{h}_{t,2}$ is stochastically dominated by an unconditioned copy of the first curve, subject to two modifications:
  1. Logarithmic Shift: A shift of order $C t^{-1/3} \log M$ is required. The authors identify this shift as a quantitative manifestation of the entropy phenomenon discussed in the introduction.
  2. Error Term: The inequality holds up to an error term of order $\exp(-cM^2)$, which arises from the probability of large fluctuations in the partition functions.
     The inequality holds for events defined on intervals disjoint from the conditioning interval, a constraint arising from the Markovian structure used in the discrete proof.

Significance and Applications
The paper positions these inequalities as key inputs for analyzing the KPZ line ensemble and the KPZ equation. Specifically, the authors note that their results:

• Validate Frameworks in [GH22]: The inequality serves as the necessary input to prove sharp upper tail estimates for the KPZ equation and to describe the limit shape of line ensembles under upper tail conditioning. The authors clarify that their work corrects an earlier, incorrect formulation of the BK inequality for the KPZ line ensemble that was relied upon in [GH22].
• Support Convergence Results in [GHZ23]: The inequality is a crucial component in proving that the continuum directed random polymer, conditioned on an upper tail event, converges to a Brownian bridge.
• Highlight Integrability: The paper underscores the fundamental role of integrability in establishing correlation inequalities for positive-temperature models, contrasting this with the zero-temperature case where such inequalities hold for general i.i.d. distributions.

The authors maintain a modest stance regarding the generality of their results, noting that while generalizations to more points or higher-indexed curves are plausible (Section 5), they require technical ingredients (such as convergence of multi-path partition functions with mixed boundary conditions) that are not currently available in the literature. The work is presented as a self-contained resolution to a specific open problem regarding the validity and form of BK inequalities in the positive-temperature KPZ class.