Neural Spectral Bias and Conformal Correlators I: Introduction and Applications

This paper demonstrates that simple feed-forward neural networks can accurately reconstruct diverse conformal field theory correlators using minimal data—specifically crossing symmetry, a single scaling dimension, and one anchor point—by leveraging the spectral bias of gradient-based training to favor the smooth functions characteristic of physical correlators.

The Gist

Imagine you are trying to solve a massive, complex jigsaw puzzle. The picture on the box is a beautiful, intricate landscape (representing the laws of physics in a specific universe, known as a Conformal Field Theory or CFT).

Usually, to solve this puzzle, you need thousands of pieces and a huge amount of time. You need to know the exact shape of every single piece (the "spectrum" of particles) and how they fit together perfectly. This is the traditional way physicists have tried to understand these theories for decades.

The Breakthrough: A Magic Guessing Machine

This paper introduces a surprisingly simple trick. Instead of needing thousands of pieces, the authors show that a Neural Network (a type of computer program that learns like a brain) can solve the puzzle using only three tiny clues:

1. The Size of the Main Piece: The "weight" or scaling dimension of the main particle involved.
2. The Gap: The size of the smallest piece that isn't the main one (a "gap" in the spectrum).
3. One Anchor Point: Just one single number telling the network what the picture looks like at one specific spot in the middle of the puzzle.

With just these three tiny inputs, the neural network is asked to draw the entire picture.

The Secret Ingredient: "Spectral Bias"

Here is the magic part. Mathematically, there are infinite ways to draw a picture that fits those three clues. Most of those drawings would look like static, scribbles, or chaotic noise. They wouldn't be the real physics.

So, why does the computer pick the right picture every time?

The authors discovered that neural networks have a built-in personality quirk called "Spectral Bias." Think of it like this:

• If you ask a human to draw a curve based on a few points, they might draw a jagged, messy line.
• But a neural network, when trained to minimize error, naturally prefers smooth, gentle, flowing curves. It hates jagged, high-frequency noise.

The authors realized that real physical laws are incredibly smooth. Nature doesn't like jagged, chaotic scribbles; it prefers elegant, flowing functions.

Because the neural network is "biased" to find the smoothest possible solution, and because the actual physical universe is the smoothest possible solution that fits the rules, the computer accidentally stumbles upon the correct physics. It's as if the computer has a "magnetic attraction" to the truth because the truth is the smoothest option.

The "Smoothness" Test

To prove this, the authors didn't just trust the computer. They acted like detectives. They took the real physical pictures and compared them to "fake" pictures that also fit the three clues but were slightly jagged or bumpy.

They used mathematical tools (like measuring the "curvature" of the lines or how fast the colors change) to prove that the real physical pictures were indeed the smoothest ones in the entire universe of possibilities. The neural network wasn't just guessing; it was following a hidden rule of the universe: Nature loves smoothness.

What Did They Solve?

They tested this "magic guessing machine" on a huge variety of puzzles:

• Simple puzzles: Basic free particles.
• Complex puzzles: The famous Ising Model (which describes how magnets work and how materials change from solid to liquid).
• Hot puzzles: How things behave at high temperatures.
• 4D puzzles: Even theories related to the Standard Model of particle physics and string theory (N=4 Super Yang-Mills).

In almost every case, the neural network, trained in just a few minutes on a standard laptop, reconstructed the complex physics with 99% accuracy.

The Big Picture

This paper suggests a new way to do physics. Instead of trying to calculate every single particle interaction from scratch, we can use the fact that physics is smooth to let computers "hallucinate" the correct answer, provided we give them a few anchor points and the rule that "the answer must be smooth."

It's like giving a child a few dots on a page and asking them to connect them. If you tell them, "Draw the smoothest line possible," they will almost always draw the beautiful curve you intended, rather than a jagged mess. The authors found that the universe is that beautiful curve, and neural networks are the perfect tools to find it.

Technical Summary

1. Problem Statement

The central challenge addressed is the reconstruction of correlation functions in Conformal Field Theories (CFTs). While the conformal bootstrap program successfully derives rigorous bounds on CFT data (scaling dimensions and OPE coefficients) using crossing symmetry and unitarity, it typically does not yield the full functional form of correlation functions $G(z, \bar{z})$ across the entire kinematic domain.

Specifically, the authors focus on 1D CFT correlators (restricted to the real line, where $z = \bar{z}$). The problem is formulated as follows:

• Input: Minimal data consisting of:
  1. The scaling dimension of the external operator ($\Delta_\phi$).
  2. The scaling dimension of the leading non-trivial operator in the OPE (the "gap," $\delta$).
  3. A single numerical value of the correlator at a specific "anchor point" $z_0$ (an "anchor" condition).
• Constraint: The function must satisfy the crossing symmetry equation:
  $$G(z) = \left( \frac{z}{1-z} \right)^{2\Delta_\phi} G(1-z)$$
• The Obstacle: This setup is severely under-determined. There exists an infinite-dimensional family of functions satisfying the crossing equation and the gap/anchor constraints. Most of these are unphysical. The goal is to identify the unique physical CFT correlator within this vast solution space using only the minimal inputs.

2. Methodology

The authors propose a novel computational strategy using Physics-Informed Neural Networks (PINNs) to solve this inverse problem.

A. Neural Network Architecture

• Parametrization: The unknown part of the correlator, $H(z)$, is represented by a lightweight Multi-Layer Perceptron (MLP) with 2–3 hidden layers and 64–128 neurons.
• Ansatz: To enforce physical constraints, the network output is dressed with prefactors:
  $$H(z) = z^\delta (1-z)^{-\delta'} \cdot \text{NN}_\theta(z)$$
  • $z^\delta$ enforces the correct small-$z$ asymptotic behavior (the gap).
  • $(1-z)^{-\delta'}$ (where $\delta' = 2\Delta_\phi$) enforces the crossing symmetry behavior near $z=1$.
  • Smooth activation functions (Tanh or GELU) are used exclusively to ensure the output is smooth.
• Loss Function: The training minimizes a loss function composed of:
  1. Crossing Loss: The mean relative squared error of the crossing equation evaluated on a grid of points in $(0, 1)$.
  2. Anchor Loss: The squared difference between the network's prediction at $z_0$ and the prescribed anchor value.

B. The Core Mechanism: Spectral Bias

The paper argues that the success of this method relies on the Spectral Bias (or Frequency Principle) of gradient-based neural network training.

• Phenomenon: Neural networks trained via gradient descent preferentially learn low-frequency, smooth components of a function before high-frequency, oscillatory ones.
• Hypothesis: Physical CFT correlators possess a unique "smoothness" property that distinguishes them from generic crossing-symmetric solutions. The NN's implicit bias toward smoothness acts as a selection mechanism, naturally guiding the optimization toward the physical solution without needing explicit regularization or complex constraints.

3. Key Contributions

A. Demonstration of Minimal-Data Reconstruction

The authors demonstrate that the anchored NN approach can reconstruct target physical correlators with percent-level accuracy (often <1% relative error) across a wide variety of theories, including:

• Generalised Free Fields (GFF): Bosonic and fermionic cases.
• AdS$_2$ Witten Diagrams: Contact and one-loop (bubble) diagrams, which involve logarithmic and polylogarithmic dependencies.
• 2D Minimal Models: Both unitary (e.g., Ising, Tricritical Ising) and non-unitary (Lee-Yang) models.
• 3D Ising Model: A genuinely predictive application where exact analytic results are unknown. The NN predictions are compared against independent bootstrap data and fuzzy-sphere simulations.
• Wilson-Fisher Fixed Points: Perturbative $\epsilon$-expansion results in $4-\epsilon$ dimensions.
• 4D $\mathcal{N}=4$ SYM: Half-BPS operator correlators in the large-$c$ limit.
• Thermal Correlators: Finite-temperature two-point functions at zero spatial separation, satisfying the KMS condition (which reduces to the crossing equation).

B. Theoretical Analysis of Smoothness

To ground the empirical success, the authors perform an independent analysis of correlator smoothness using three diagnostic tools:

1. Fractional Sobolev Semi-norms: Measures global regularity.
2. Chebyshev Spectral Decomposition: Analyzes the decay rate of spectral coefficients (analytic functions exhibit super-algebraic decay).
3. Curvature-Based Measures: Integrals of the squared second derivative.

• Finding: Physical CFT correlators consistently appear as the smoothest functions among families of crossing-symmetric deformations. This confirms that the NN's spectral bias is selecting the correct physical solution because it minimizes a smoothness functional.

C. Extension Beyond the Line

The authors extend the method to 2D kinematics (the complex plane) by training on concentric circles centered at the crossing-symmetric point $z = \bar{z} = 1/2$. This reduces the 2D crossing constraint to a 1D problem on the circle, successfully reconstructing correlators without additional input beyond the real-line anchor.

D. Alternative Approach: Chebyshev-Tikhonov

The paper explores a non-NN alternative using Chebyshev polynomial expansions with Tikhonov regularization (penalizing high-frequency coefficients). While this can also fit correlators, it requires manual tuning of regularization parameters. The NN approach is highlighted as superior because the smoothness bias is implicit and automatic, requiring no case-specific fine-tuning.

4. Key Results

• Accuracy: In all tested cases, the NN predictions match exact analytic results or high-precision numerical data (from bootstrap/fuzzy-sphere) with relative errors typically below 1%.
• Robustness: The method works for unitary and non-unitary theories, bosonic and fermionic operators, and both vacuum and thermal correlators.
• Predictive Power: For the 3D Ising model (where no closed-form solution exists), the NN provided new numerical data for the $\langle \epsilon \epsilon \epsilon \epsilon \rangle$ correlator that agreed closely with independent fuzzy-sphere simulations ($3.987 \pm 0.027$ vs. $4.04$).
• Efficiency: The method converges in minutes on a standard laptop using lightweight networks.

5. Significance and Outlook

• New Computational Paradigm: This work introduces a "minimal-data" approach to the conformal bootstrap, bypassing the need for extensive spectral data or complex optimization over infinite-dimensional spaces.
• Variational Principle: The results suggest the existence of a deep, theory-independent variational principle for 1D CFT correlators: physical correlators are the minimizers of a smoothness functional (related to Sobolev norms) among all crossing-symmetric functions with fixed gap and anchor data.
• Bridge to Machine Learning: It provides a concrete physical example where the "spectral bias" of neural networks is not a bug but a feature that aligns perfectly with the mathematical structure of quantum field theory.
• Future Directions: The authors propose using this method to extract OPE data, combine it with dispersion relations, and apply it to more complex systems (mixed correlators, spinning operators, and higher-dimensional modular bootstrap).

In summary, the paper establishes that simple neural networks, guided by the principle of smoothness, can solve the conformal bootstrap problem for 1D correlators using only minimal input data, offering a powerful, non-perturbative tool for exploring the landscape of Conformal Field Theories.