Critical curve of two-matrix models $ABBA$, $A\{B,A\}B$ and $ABAB$, Part I: Monte Carlo

This paper presents Monte Carlo estimates for the critical curves (boundaries of maximal convergence domains) in the $(h,g)$-plane for three specific two-matrix models with interaction terms $ABBA$, $A\{B,A\}B$, and $ABAB$, comparing these numerical results with exact solutions and functional renormalization group phase diagrams.

The Gist

Imagine you are an architect trying to build a skyscraper, but instead of bricks and steel, your building blocks are random matrices (giant grids of numbers). You want to build a stable structure, but you have two knobs you can turn to change how these blocks interact:

1. Knob $g$ (The "Stiffness" Knob): Controls how much the blocks want to stay in their own shape.
2. Knob $h$ (The "Mixing" Knob): Controls how much the two different types of blocks (let's call them Type A and Type B) want to dance together.

The paper by Carlos I. Pérez Sánchez is essentially a map of stability. It asks: "If I turn these knobs to specific settings, will my skyscraper stand tall, or will it collapse into chaos?"

Here is the breakdown of the paper's journey, using simple analogies.

1. The Three Different Dance Floors

The author isn't just looking at one way the blocks can dance. He is testing three different choreographies (mathematical formulas) for how Type A and Type B interact:

• The "ABBA" Dance: They go A, then B, then B, then A. (Like a palindrome).
• The "ABAB" Dance: They go A, then B, then A, then B. (Like a zipper).
• The "Mix" Dance: A combination of the two, controlled by a slider called $q$.

The goal is to find the Critical Curve. Imagine a map where the safe, stable zone is a green island, and the chaotic, collapsing zone is a red ocean. The "Critical Curve" is the shoreline. If you step just one inch over the line, your building falls apart.

2. The Problem: We Can't Solve the Math

In the world of physics, some problems are easy to solve with a pen and paper (like a single block dancing). But when you have two blocks dancing together in complex ways, the math becomes a tangled knot that no human (or even current supercomputers) can untangle with a formula.

The author notes that for one specific dance (the ABAB dance), mathematicians solved it years ago. But for the others, we are flying blind. We don't know where the shoreline is.

3. The Solution: The "Virtual Architect" (Monte Carlo)

Since we can't solve the equation, the author builds a Virtual Architect (a computer simulation using a method called Hamiltonian Monte Carlo).

• The Simulation: The computer starts with a random set of blocks. It tries to build the structure.
• The Test: It checks if the structure holds together.
  • If the numbers stay calm and finite, the computer says: "Green Light! This is safe."
  • If the numbers explode to infinity (the building collapses), the computer says: "Red Light! This is dangerous."

4. The Smart Search Strategy

The author realized that testing every single point on the map is too slow (it would take forever). So, he invented a Smart Search Strategy:

• The Dipole Hunt: Instead of checking random spots, the computer looks for a "Dipole." This is a pair of points: one Green (Safe) and one Red (Unsafe) that are very close to each other.
• The Midpoint: Once it finds a Green point and a Red point right next to each other, it knows the "shoreline" (the Critical Curve) must be somewhere in between. It marks that spot.
• The Radial & Angular Sweep: The computer sweeps outwards from the center like a radar, and also spins around in circles, hunting for these Green/Red pairs to draw the shoreline as accurately as possible.

5. The Big Discoveries

After running millions of these virtual simulations, the author found some surprising things:

• The "ABBA" and "Mix" dances are twins: The shoreline for the "ABBA" dance and the "Mix" dance (where $q=0.5$) look almost identical. It's as if they are dancing to the same rhythm.
• The "ABAB" dance is the odd one out: The "ABAB" dance has a very different shoreline. It behaves differently when you turn the knobs in the negative direction.
• The "Infinite" Limit: The author figured out what happens when you turn the "Mixing" knob ($h$) to infinity. The shoreline becomes a straight line at a specific angle. He measured these angles precisely for all three dances.

6. Why Does This Matter?

You might ask, "Who cares about random number grids?"

• Quantum Gravity: These matrices are used to model the fabric of spacetime itself. Understanding where the "stable" zones are helps physicists understand how the universe might have formed or how black holes behave.
• New Tools: The author didn't just find the map; he built a better compass. His computer code can now be used to map any complex interaction between two types of variables, not just these specific ones.

Summary Analogy

Imagine you are trying to find the edge of a foggy cliff.

• Old way: Walk randomly until you fall off. (Inefficient and dangerous).
• This paper's way: You have a robot that can sense the ground. It walks until it finds a safe spot, then takes a tiny step. If it falls, it knows the edge is between the last safe spot and the fall. It marks that spot, then moves to a new angle and repeats.
• Result: The author has drawn a precise map of the cliff's edge for three different types of fog, showing us exactly where the ground ends and the void begins.

This paper is a triumph of computational physics: using brute-force simulation combined with clever algorithms to solve a problem that pure math couldn't crack.

Technical Summary

1. Problem Statement

The paper investigates the critical loci (boundaries of the maximal convergence domain) for a family of two-matrix models involving two $N \times N$ Hermitian matrices, $A$ and $B$. The models are defined by the action:
$$S^{(q)}_{g,h}(A, B) = \frac{1}{2}\text{Tr}(A^2 + B^2) - \frac{g}{4}\text{Tr}(A^4 + B^4) - \frac{h}{2}\text{Tr}(A\{B, A\}_q B)$$
where $\{B, A\}_q = qBA + (1-q)AB$ is a $q$-deformed commutator/anticommutator, with $q \in [0, 1]$.

The study focuses on three specific cases:

• $q=0$ (ABBA model): Interaction term $\text{Tr}(ABBA)$.
• $q=1/2$ (A{B, A}B model): Interaction term $\text{Tr}(A\{B, A\}B) = \frac{1}{2}\text{Tr}(ABAB + ABBA)$.
• $q=1$ (ABAB model): Interaction term $\text{Tr}(ABAB)$.

The Core Challenge: Unlike one-matrix models, which are solvable via topological recursion, general two-matrix models with arbitrary potentials lack a complete analytic solution. The ABAB model ($q=1$) is a rare exception solved analytically by Kazakov and Zinn-Justin. The goal is to numerically determine the phase boundaries (critical curves) in the $(g, h)$ coupling plane for these models, specifically where the partition function $Z_N(g, h)$ ceases to exist (diverges).

2. Methodology

The author employs Hamiltonian Markov Chain Monte Carlo (HMC) simulations to estimate the critical curves.

• Simulation Setup:

  • Matrix Size: $N=32$ is chosen as the optimal size to balance computational cost with the need to approximate the large-$N$ limit (where corrections are of order $N^{-2}$).
  • Iterations: Markov chains are run for $n = 20,000$ iterations to ensure thermalization and detect divergence accurately.
  • Initialization: Chains start from $X_1 = 0_N$ (zero matrix) to avoid premature rejection of proposals that might occur if starting from a Schwinger-Dyson solution.

• Dynamics (Leapfrog Integration):

  • The algorithm introduces conjugate momenta $P$ and evolves the system using Hamiltonian dynamics: $\dot{P} = -N\nabla_X S$ and $\dot{X} = P$.
  • The "force" term is derived from the matrix derivatives of the action.
  • A Metropolis-Hastings acceptance criterion is used to accept or reject proposed configurations based on the change in action $\Delta S$.

• Criticality Detection (The "MC" Function):

  • A boolean function MC(g, h) determines convergence.
  • Convergence (True): Expectation values remain finite, matrices retain Hermiticity, and the eigenvalue distribution remains healthy throughout the simulation.
  • Divergence (False): Expectation values explode, Hermiticity is irreversibly lost, or the simulation aborts due to numerical instability.

• Search Algorithms for the Critical Curve:
  To efficiently map the boundary between convergent (True) and divergent (False) regions without testing a dense static grid, the author developed dynamic search routines:

  1. Midpoint Search: Bisection method between a known True point and a False point to find the boundary.
  2. Radial Search: Starting from a False point, the algorithm moves radially toward the origin (where the model is known to converge) until a True point is found, creating a "dipole" (True/False pair) to define the boundary.
  3. Angular Search: Used for large couplings. Points are tested along a circular arc to find the transition from True to False, determining the asymptotic slope of the critical curve.

• Thermalization & Validation:

  • Schwinger-Dyson Equations (SDEs): The code monitors the residual of SDEs (LHS - RHS) during the simulation. The thermalization time $\tau$ is determined when the sum of squared SDE residuals drops below a threshold $\epsilon$.
  • Positivity Checks: The algorithm monitors the positivity of the moment matrix (a condition required for the existence of the measure), ensuring physical consistency.

3. Key Contributions

1. Numerical Determination of Critical Curves: The paper provides the first Monte Carlo estimates of the critical boundaries for the $q=0$ (ABBA) and $q=1/2$ models, which lack analytic solutions.
2. Validation of Analytic Results: For the $q=1$ (ABAB) model, the numerical results perfectly match the known analytic solution by Kazakov-Zinn-Justin, validating the Monte Carlo approach.
3. Asymptotic Analysis: The study characterizes the behavior of the critical curves as $h \to \pm \infty$, calculating the slopes ($\lambda_\pm$) and angles ($\Theta_\pm$) for all three models.
4. Dynamic Search Algorithms: The introduction of radial and angular search methods significantly reduces computational cost compared to static lattice testing, allowing for the mapping of complex phase diagrams.
5. Insight into $q$-Deformation: The results suggest a qualitative difference between models with $q \le 1/2$ and $q=1$. Specifically, the $q=1$ model exhibits a symmetry $h \to -h$ in its critical curve that is broken or altered in the $q=1/2$ and $q=0$ cases.

4. Key Results

The critical curves were mapped in the $(g, h)$ plane. Key findings include:

• Critical Points on Axes:

  • At $h=0$, all models share the critical point $g_c = 1/12$ (pure gravity limit).
  • At $g=0$, the critical $h$ values differ:
    • $q=1$ (ABAB): $h_c = 2/9$ (matches analytic solution).
    • $q=1/2$ and $q=0$: The critical $h$ values are lower than $2/9$, consistent with the inequality derived from operator positivity.

• Asymptotic Slopes ($h \to \pm \infty$):
  The paper defines $\lambda_\pm(q)$ as the slope of the critical line in the limit of large $h$.

  • $q=1$ (ABAB): $\lambda_+ \approx -0.97$, $\lambda_- \approx +0.98$. The curve is symmetric ($h \to -h$).
  • $q=1/2$ (A{B, A}B): $\lambda_+ \approx -0.98$, $\lambda_- \approx 0$. The curve becomes horizontal for $h \to -\infty$.
  • $q=0$ (ABBA): $\lambda_+ \approx -0.97$, $\lambda_- \approx 0$. Similar to $q=1/2$.

• Phase Diagram Structure:

  • The $q=1$ model has a "diamond-like" shape in the positive quadrant.
  • The $q=0$ and $q=1/2$ models exhibit a "flat" asymptote for negative $h$ (slope $\approx 0$), indicating a different universality class or behavior compared to the oscillatory $q=1$ case.
  • The region of convergence for $q \le 1/2$ is strictly contained within the region for $q=1$ in the positive quadrant, consistent with theoretical bounds.

5. Significance and Future Directions

• Benchmarking: This work serves as a crucial benchmark for other numerical methods, such as the Functional Renormalization Group (FRG). The author notes that previous FRG studies of the ABAB model produced phase diagrams that lacked the correct $h \to -h$ symmetry, suggesting that the FRG truncations used were insufficient. The Monte Carlo results confirm that the $h^2$ term in the FRG beta-functions must vanish for the ABAB model.
• Universality Classes: The results suggest that the family of models splits into two classes: $q=1$ (oscillatory, symmetric) and $q \le 1/2$ (non-oscillatory, asymmetric).
• Applications: The ABAB model is linked to Causal Dynamical Triangulations (CDT). Understanding the critical curves of related models (like ABBA) could provide insights into the gluing of geometries in quantum gravity.
• Part II: This paper focuses on Monte Carlo. The author indicates that Part II will address the same models using positivity bootstraps (moment matrix methods) and further FRG analysis, aiming to provide analytic constraints and cross-validation.

In summary, this paper successfully applies advanced HMC techniques to solve a long-standing problem in random matrix theory: mapping the phase diagrams of non-solvable two-matrix models, revealing distinct critical behaviors based on the deformation parameter $q$.