Delocalization of the height function of the six-vertex model

This paper proves that the height function of the six-vertex model with parameters $a=b=1$ and $1 \le c \le 2$exhibits delocalization with logarithmic variance, thereby completing the characterization of the model's phase transition by complementing the known localization behavior for$c > 2$.

The Gist

Imagine a giant, infinite checkerboard where every square is a tiny room. Inside each room lives a number (a "height"). The rule of this game is simple: if you walk from one room to an adjacent room, the number must go up or down by exactly one. You can't jump from 5 to 7; you must go 5 to 6 to 7.

This is the Six-Vertex Model. It's a mathematical way to describe how things like ice crystals form, how water molecules arrange themselves, or how a random surface might look.

The big question the authors of this paper are asking is: If you stand in one room and look at a room far away, how different are the numbers?

The Two Worlds: Smooth vs. Rough

The authors discovered that the answer depends on a single "knob" on the machine, a parameter they call $c$.

1. The Smooth World ($c > 2$): The Localized Phase
Imagine you are painting a wall. If the paint is very thick and sticky (high $c$), it doesn't spread far. If you make a bump in the paint here, it stays a bump there. The wall stays relatively flat.
In this world, the height difference between two distant rooms stays small and bounded. No matter how far apart you are, the numbers don't drift too far apart. The surface is localized (smooth).

2. The Rough World ($1 \le c \le 2$): The Delocalized Phase
Now, imagine the paint is very watery and runny (low $c$). If you make a tiny bump, the water ripples out. A small push here causes a wave that travels far.
In this world, the height difference between two distant rooms grows as you move further apart. Specifically, it grows like the logarithm of the distance.

• If you double the distance, the difference in height doesn't double; it adds a little bit more.
• If you go 100 steps away, the height might be 2 units different.
• If you go 10,000 steps away, it might be 4 units different.
• If you go a million steps away, it might be 6 units different.

This is called delocalization. The surface is "rough" and keeps fluctuating wildly over long distances.

The Big Discovery

Before this paper, mathematicians knew about the "Smooth World" ($c > 2$) and they knew about a few special cases of the "Rough World" (like when $c=1$ or $c=\sqrt{2}$). But for the vast middle ground ($1 < c \le 2$), nobody could prove for sure if the surface was smooth or rough.

This paper proves that for the entire middle range ($1 \le c \le 2$), the surface is Rough. It fluctuates, it wanders, and it never settles down.

How Did They Solve It? (The Detective Work)

The authors used three main tools to crack the case, which they describe in the paper using fancy math, but here is the simple version:

1. The "Free Energy" Thermometer
Think of "Free Energy" as a measure of how much the system wants to move.

• In the Smooth World, the system is stiff. It takes a lot of effort to create a slope, so the "energy cost" curve is jagged (non-differentiable) at zero.
• In the Rough World, the system is flexible. The energy cost curve is smooth and round at zero.
  The authors used a known formula (from the "Bethe Ansatz," a famous physics trick) to show that for $c \le 2$, this energy curve is perfectly smooth. This hinted that the surface should be rough.

2. The "Fence" Strategy (RSW Theory)
To prove the surface is rough, they had to show that "loops" of high numbers can form easily.
Imagine trying to build a fence around a garden. If you can easily build a fence of height $k$, you can also build a fence of height $2k$ (maybe not easily, but with some probability).
They used a technique called RSW (Russo-Seymour-Welsh). Think of it like a game of "connect the dots." If you can cross a small square with a high number, you can combine many small squares to cross a huge rectangle. They proved that even if you start with a tiny chance of a high number, you can build a "circuit" (a loop) of high numbers around a large area with a decent probability.

3. The "Height vs. Absolute Value" Trick
This is the cleverest part. The math gets messy because the numbers can be positive or negative (up or down).
The authors realized that if you look at the absolute value (ignoring the sign, just looking at how "tall" the bump is), the rules become much simpler and friendlier. They proved that if you can build a loop of "tall" bumps (regardless of whether they are up or down), the surface must be rough. They used a connection to the Ising Model (a famous model of magnets) to show that the "tallness" behaves like a magnetic field that wants to align, making it easier to prove the loops exist.

The Analogy: The Drunkard's Walk

Imagine a drunkard walking on a grid.

• Smooth World ($c > 2$): The drunkard is tied to a leash. He wanders a bit, but he never gets very far from the starting point. His distance from home stays small.
• Rough World ($1 \le c \le 2$): The drunkard is untied. He wanders aimlessly. Over time, his distance from home grows. It doesn't grow linearly (he doesn't walk in a straight line), but it grows slowly, like the square root of time (or logarithmically in this specific 2D case). He is delocalized.

Why Does This Matter?

This isn't just about math puzzles.

1. Ice and Water: The six-vertex model was originally invented to understand why ice has a specific entropy (disorder) at absolute zero. Understanding the "roughness" helps us understand the microscopic structure of ice.
2. Universality: This result confirms a deep idea in physics: that many different systems (magnets, fluids, ice) behave the same way near critical points. This paper proves that the "Rough" behavior is the standard rule for this whole family of models when the parameters are right.
3. The Gaussian Free Field: The authors suggest that in this rough phase, the surface looks like a Gaussian Free Field (GFF). Think of this as a perfectly random, wiggly surface that nature loves to use. It's the "standard model" of random surfaces in physics.

Summary

The paper solves a decades-old mystery: When the parameters of the six-vertex model are set to $1 \le c \le 2$, the surface is not flat. It is a wild, fluctuating, rough landscape that grows logarithmically with distance. They proved this by combining deep energy calculations with clever geometric arguments about building loops and fences, finally connecting the dots between the "smooth" and "rough" phases of this mathematical universe.

Technical Summary

Here is a detailed technical summary of the paper "Delocalization of the height function of the six-vertex model" by Hugo Duminil-Copin, Alex Karrila, Ioan Manolescu, and Mendes Oulamara.

1. Problem Statement and Motivation

The paper addresses the fundamental question of localization versus delocalization for the height function of the six-vertex model on the two-dimensional lattice.

• The Model: The six-vertex model (also known as the square ice model) is a statistical mechanics model where arrows on the edges of a lattice satisfy the "ice rule" (two incoming, two outgoing arrows at every vertex). It can be mapped to a random height function $h$ defined on the faces of the lattice, where $|h(x) - h(y)| = 1$ for adjacent faces.
• The Regime: The authors focus on the parameter range $a=b=1$ and $c \ge 1$. This corresponds to the antiferroelectric phase (specifically Rys' model of hydrogen-bonded ferroelectrics).
• The Dichotomy:
  • Localized (Smooth): The variance of the height difference, $\text{Var}(h(x) - h(y))$, remains bounded as the distance $d(x,y) \to \infty$.
  • Delocalized (Rough): The variance grows to infinity (typically logarithmically) as $d(x,y) \to \infty$.
• Previous Knowledge: It was known that for $c > 2$, the model is localized. For specific values $c=1$ (square ice), $c=\sqrt{2}$ (free fermion point), and $c=2$, delocalization was established or conjectured. However, a rigorous proof for the entire range $1 \le c \le 2$ was missing.
• Goal: To prove that for $1 \le c \le 2$, the height function is **delocalized with logarithmic variance**, i.e.,$c \log d(x,y) \le \text{Var}(h(x) - h(y)) \le C \log d(x,y)$.

2. Methodology and Core Ideas

The proof is non-perturbative and relies on a combination of probabilistic techniques, combinatorial properties of the model, and exact results from integrable systems. The methodology is structured around three main pillars:

A. Spatial Markov Property and Monotonicity

The authors utilize the Spatial Markov Property (SMP) and FKG (Fortuin-Kasteleyn-Ginibre) inequalities.

• Crucially, for $c \ge 1$, the model exhibits positive association (FKG) not only for the height function $h$ but also for its absolute value $|h|$. This allows for the comparison of boundary conditions (CBC) and the "pushing/pulling" of boundaries, which are essential for renormalization arguments.
• They establish that the model is invariant under vertical and diagonal reflections when $a=b$, which is critical for the geometric arguments.

B. RSW Theory (Russo-Seymour-Welsh)

The paper develops an RSW-type theory for the level sets of the height function.

• Challenge: Standard RSW theory for percolation compares crossing probabilities of rectangles of different aspect ratios. Here, the authors must handle height constraints.
• Innovation: They prove that the probability of a "high" crossing (height $\ge ck$) in a long domain can be bounded from below by the probability of a "lower" crossing (height $\ge k$) in a shorter domain.
• Theorem 3.1: Establishes that crossing probabilities are uniformly bounded away from zero for annuli, provided the boundary conditions are "flat" (bounded). This is achieved by constructing circuits from smaller crossings and using the FKG property for $|h|$.

C. Free Energy and Differentiability

The most significant theoretical input comes from the free energy of the cylindrical six-vertex model.

• Theorem 1.6 (from [18]): The free energy $f_c(\alpha)$ of the model on a cylinder with an unbalance $\alpha$ (difference between up and down arrows) is differentiable at $\alpha=0$ if and only if $c \le 2$. For $c > 2$, it is non-differentiable (indicating a phase transition).
• The Bridge (Theorem 1.7): The authors prove a novel connection: the differentiability of the free energy at zero implies a lower bound on the probability of annulus circuits. Specifically, they relate the probability of a circuit of height $k$ to the exponential of the free energy difference:
  $$P(\text{circuit}) \gtrsim \exp\left[ C r^2 (f_c(k/\eta r) - f_c(0)) \right]$$
  Since $f_c$ is differentiable (quadratic behavior near 0) for $c \le 2$, this probability does not decay exponentially fast, allowing for the construction of large circuits.

3. Key Contributions and Results

The paper provides the first complete description of the localization/delocalization transition for this parameter range.

Main Theorems

1. Theorem 1.1 (Delocalized Phase): For $1 \le c \le 2$, on a large torus$T_N$, the variance of the height difference satisfies:
   $$c \log d(x,y) \le \mathbb{E}[(h(x) - h(y))^2] \le C \log d(x,y)$$
   This confirms the logarithmic divergence characteristic of the Gaussian Free Field (GFF).

2. Theorem 1.4 (Uniform Circuit Probabilities): For $c \in [1, 2]$, the probability of finding a circuit of height $\ge k$ surrounding a box of size $n$ in an annulus of size $2n$is uniformly bounded from below by a positive constant, independent of the scale$n$ and the domain, provided boundary conditions are flat.

3. Corollary 1.5 (Planar Domains): The logarithmic variance bounds hold for arbitrary finite planar domains with admissible boundary conditions.

Technical Innovations

• From Free Energy to Circuits: The derivation of Theorem 1.7 is the paper's main innovation. It translates the analytic property of the free energy (differentiability) into a geometric probabilistic statement (existence of high-level circuits).
• Handling Absolute Values: The proof heavily relies on the FKG property for $|h|$, which is non-trivial and specific to the $c \ge 1$ regime. This allows the authors to control the "fences" (low height barriers) that separate high height regions.
• Renormalization without Loss: Unlike standard RSW proofs where height might decrease during renormalization (making it hard to iterate), the authors use the free energy input to show that the renormalization step is "superfluous" in a sense; the differentiability ensures the probability of circuits remains high enough without needing to track a decaying height parameter.

4. Significance and Implications

• Completing the Phase Diagram: The paper resolves the behavior of the six-vertex model for the entire range $1 \le c \le 2$. It confirms that the transition from delocalized to localized occurs exactly at$c=2$ (where the free energy becomes non-differentiable).
• Gaussian Free Field (GFF) Convergence: The logarithmic variance is a necessary condition for the height function to converge to the GFF in the scaling limit. While the paper proves the variance scaling, it notes that full convergence to the GFF is known only for $c=\sqrt{2}$ and $c=2$, but the results here strongly support the conjecture for the whole range.
• Methodological Impact: The technique of linking the differentiability of the free energy (an integrable systems result) to geometric circuit probabilities (a probabilistic result) provides a powerful new tool for studying phase transitions in lattice models. It bypasses the need for explicit correlation function calculations, which are often intractable.
• Universality: The results suggest that the delocalization behavior is robust across the parameter range $1 \le c \le 2$, reinforcing the idea that these models belong to the same universality class as the GFF.

Summary of the Proof Logic

1. Integrable Input: Use the known result that free energy $f_c(\alpha)$ is differentiable at 0 for $c \le 2$.
2. Probabilistic Bridge: Prove that differentiability implies a lower bound on the probability of "ridge" events (circuits of high height) on a cylinder (Theorem 1.7).
3. Geometric Construction: Use RSW-type arguments (Theorem 3.1) to extend these cylinder results to planar domains and annuli, establishing that high circuits exist with non-vanishing probability.
4. Variance Bounds:
   • Lower Bound: Use the existence of high circuits to show that the height must fluctuate significantly (logarithmic growth).
   • Upper Bound: Use the FKG property and the existence of low-height "fences" to show that the height cannot fluctuate more than logarithmically.

This work represents a major step forward in the rigorous understanding of two-dimensional integrable lattice models and their connection to random surface theory.