Joint distributions of eigenvectors of symmetric random tensors

This paper employs quantum field theoretical methods to compute the joint distributions of arbitrary numbers of eigenvectors for real and complex symmetric random tensors, deriving their random matrix representations and large-dimension asymptotics to demonstrate a universal behavior across tensor geometries that extends previous findings on mean distributions.

The Gist

The Big Picture: Finding Patterns in Chaos

Imagine you have a giant, multi-dimensional puzzle. In the world of math and physics, these puzzles are called tensors. While a matrix is a 2D grid of numbers (like a spreadsheet), a tensor is a 3D, 4D, or even higher-dimensional block of numbers.

These tensors are everywhere in modern science, from understanding how AI learns to modeling the gravity of black holes. However, solving these puzzles is incredibly hard. If you try to find all the "solutions" (called eigenvectors) for a specific, random puzzle, there are so many of them that the number explodes exponentially. It's like trying to count every single grain of sand on a beach while the beach keeps growing.

Because counting them all is impossible, scientists study random tensors. Instead of looking at one specific, messy puzzle, they look at the average behavior of millions of random puzzles. This paper takes that idea a step further.

The Problem: Looking at One vs. Looking at a Group

Previous studies were like looking at a crowd of people and asking, "What is the average height?" They found the mean distribution (the average shape of the solutions).

This paper asks a more complex question: "If I pick two, three, or ten people from this crowd, how are they related to each other?"

In math terms, the authors are studying the joint distributions of eigenvectors. They want to know the probability of finding specific eigenvectors together. Do they tend to cluster? Do they avoid each other? Are they independent?

The Method: A Quantum Field Theory "Magic Trick"

The authors use a sophisticated tool from theoretical physics called Quantum Field Theory (QFT). To understand this, imagine you are trying to predict the weather. Instead of simulating every single air molecule (which is too hard), you use a "field" model that treats the air as a continuous fluid.

The authors use a similar "field" approach to handle the massive number of solutions:

1. The Setup: They treat the random tensor like a field of energy.
2. The Transformation: They use a mathematical "magic trick" (involving bosons and fermions, which are just types of variables in this context) to turn the impossible problem of counting solutions into a problem of calculating the properties of a Random Matrix.
3. The Result: They successfully translate the complex tensor problem into a simpler "Random Matrix" problem. This is like turning a chaotic storm into a predictable wave pattern.

The Key Discovery: A Universal Shape

The most exciting finding in the paper is what happens when the dimensions get very large (the "Large-N limit").

Imagine you have different types of random puzzles (some made of real numbers, some of complex numbers). You might expect them to behave very differently. However, the authors found that when the puzzles get huge, the way their solutions relate to each other converges into a single, universal shape.

They discovered that the joint distribution of these eigenvectors can be described by one common function based on the "geometry" of the tensor.

• The Analogy: Imagine you have a bag of different colored marbles (real tensors) and a bag of glass marbles (complex tensors). If you shake them gently, they look different. But if you shake them violently (large dimensions), they all settle into the exact same pattern of stacking. The paper found the mathematical formula for that universal stacking pattern.

The Verification: Checking the Work

You might wonder, "Is this just fancy math, or does it actually work?"

The authors didn't just stop at the theory. They performed Monte Carlo simulations.

• The Test: They used computers to generate thousands of random tensors and explicitly solved for their eigenvectors (the "hard way").
• The Comparison: They compared these computer results with their new "Random Matrix" formulas.
• The Outcome: The results matched perfectly. The computer data (dots) lined up exactly with the theoretical curves (lines), even for very large systems. This confirms that their "magic trick" of turning tensors into matrices works.

Summary

In simple terms, this paper:

1. Solved a hard problem: It figured out how to calculate the probability of finding multiple solutions together in random, multi-dimensional puzzles.
2. Found a shortcut: It showed that you can solve this by converting the puzzle into a simpler matrix problem.
3. Discovered a rule: It proved that for very large systems, all these different types of puzzles follow the exact same geometric rule for how their solutions relate to one another.
4. Proved it: It used computer simulations to verify that the math is correct.

The paper essentially provides a new, efficient map for navigating the chaotic landscape of high-dimensional random systems, showing that even in chaos, there is a hidden, universal order.

Technical Summary

Technical Summary: Joint Distributions of Eigenvectors of Symmetric Random Tensors

Problem Statement
While random matrix theory has been extensively applied to linear problems, the properties of eigenvalues and eigenvectors for random tensors remain less understood due to their intrinsic complexities. Specifically, the number of eigenvalues for a tensor is generally exponential in the tensor's dimension, and computing all eigenvectors for an arbitrary tensor is computationally intractable. Although previous studies have established the mean distributions of eigenvalues and eigenvectors for real and complex symmetric random tensors, the joint distributions of arbitrary numbers of eigenvectors—which describe mutual probabilistic correlations—had not been fully derived using field-theoretical methods. This paper addresses the computation of these joint distributions for both real and complex symmetric random tensors.

Methodology
The authors employ quantum field theoretical (QFT) methods, previously utilized for mean distributions, to derive the joint distributions. The core approach involves:

1. Field Theoretical Formulation: The probability distribution of eigenvectors is expressed using delta functions and Jacobian determinants. To handle the Jacobian determinants and delta functions, the authors introduce superfields (combining bosonic and fermionic variables) and utilize supersymmetry.
2. Integration over Tensor and Auxiliary Variables: The integration over the random tensor components (Gaussian distributed) and auxiliary Lagrange multipliers is performed explicitly. This decouples the tensor from the eigenvector variables, leaving an effective action involving superfields.
3. Random Matrix Representation: The remaining interaction terms in the superfield action are linearized using Hubbard-Stratonovich transformations, introducing auxiliary matrix variables. This converts the problem into a random matrix model involving determinants of matrices dependent on the eigenvectors.
4. Large-$N$ Asymptotics: To analyze the behavior in the limit of large tensor dimensions ($N \to \infty$), the authors employ Schwinger-Dyson equations. They assume that supersymmetry and orthogonal (or unitary) symmetries are preserved in the large-$N$ limit, allowing the calculation of effective actions via two-field correlation functions.
5. Numerical Crosschecks: The analytical results are validated through Monte Carlo (MC) simulations. Two methods are compared: (i) direct computation of eigenvectors from randomly generated tensors (feasible only for small $N$), and (ii) computation of the derived random matrix representations (feasible for large $N$).

Key Contributions and Results

• Derivation of Joint Distributions: The paper derives the joint distributions for an arbitrary number ($L$) of eigenvectors for both real and complex symmetric random tensors of order $p \ge 3$.
• Random Matrix Representations:
  • For the real case, the joint distribution is expressed as an integral over real symmetric matrices. The result involves correlation functions of matrix eigenvalues, specifically utilizing skew orthogonal polynomials for explicit evaluation in the mean case ($L=1$).
  • For the complex case, the joint distribution is expressed as an integral over complex symmetric matrices. The solution utilizes the Autonne-Takagi factorization and Laguerre ensembles for the mean case.
• Large-$N$ Asymptotics:
  • The authors obtain the large-$N$ asymptotic forms of the joint distributions.
  • A significant finding is that the large-$N$ asymptotics for both real and complex cases can be expressed by a common function of tensor geometric quantities (specifically, metrics constructed from the eigenvectors).
  • The result for the real case is given by Equation (79), and for the complex case by Equation (124). The complex case's exponent is exactly twice that of the real case, consistent with the doubling of degrees of freedom.
  • The function governing the asymptotics, $h_p[x]$, is identified as a universal function previously found for mean distributions, suggesting an extension of universality to joint distributions.
• Numerical Verification:
  • Monte Carlo simulations confirm the validity of the random matrix representations for small $N$ (e.g., $N=4$) by comparing them with direct tensor eigenvector solutions.
  • Comparisons between the random matrix representations and the large-$N$ asymptotics for larger $N$ (up to $N=400$) show convincing agreement, supporting the derived asymptotic formulas.

Significance and Claims
The paper claims to extend the understanding of random tensor properties from mean distributions to joint distributions, which are crucial for studying mutual correlations and energy landscape structures. The primary significance lies in the demonstration that the large-dimension asymptotics of these joint distributions are governed by a single universal function of tensor geometric quantities. This finding suggests a potential universality across different types of random tensors, extending the universality previously established for mean distributions.

The authors modestly note that while the random matrix representations provide explicit analytic expressions primarily for mean distributions, they offer a numerically efficient method for computing joint distributions, which is superior to solving non-linear polynomial equations for large tensors. The paper concludes that these representations allow for the application of matrix model techniques (such as large deviation theory) to tensor eigenproblems and that the observed universality warrants future investigation into other types of random tensors.