Neural Networks, Dispersion Relations and the Thermal Bootstrap

This paper reviews a novel conformal bootstrap framework that replaces traditional positivity constraints with dispersion relations and neural networks to analyze scalar thermal two-point functions, demonstrating its numerical stability and application to Generalized Free Fields and 4d holographic CFTs.

The Gist

Imagine you are trying to solve a massive, infinite jigsaw puzzle. In the world of theoretical physics, this puzzle represents the rules that govern how particles interact in a "Conformal Field Theory" (CFT). Usually, physicists solve these puzzles by looking for pieces that must be positive numbers (like weights on a scale), which helps them eliminate wrong answers quickly.

However, this paper tackles a specific, trickier puzzle: thermal physics (how these theories behave at a hot temperature). In this hot environment, the "positive number" rule disappears, and the puzzle becomes a chaotic mess of infinite pieces with no obvious way to sort them.

Here is how the authors, Vasilis Niarchos and Constantinos Papageorgakis, propose to solve it, using a mix of old-school math and modern Artificial Intelligence.

1. The Problem: The Infinite Tower

In these hot theories, the puzzle involves an infinite "tower" of heavy, high-energy particles.

• The Old Way: Physicists usually try to ignore the top of the tower (the heaviest particles) and just guess what they look like. This is like trying to finish a 10,000-piece puzzle by only looking at the bottom 100 pieces and hoping the rest fits. It often leads to errors.
• The New Approach: The authors say, "Let's not guess. Let's describe the entire infinite tower mathematically."

2. The Toolkit: Dispersion Relations and Neural Networks

To handle the infinite tower without making bad guesses, they use two main tools:

• Dispersion Relations (The "Shadow" Method): Imagine you have a complex 3D object, but you can only see its shadow on a wall. The authors use a mathematical trick called a "dispersion relation" to reconstruct the whole object by analyzing its "shadow" (mathematical discontinuities). This allows them to package the infinite heavy particles into a single, manageable mathematical term.
• Neural Networks (The "Shape-Shifter"): For the remaining particles that are too light to be in the "shadow" but too heavy to list individually, they use a Neural Network. Think of this as a digital clay model. Instead of listing every single particle, they give the AI a lump of clay and tell it, "Mold this clay to fit the rules of the puzzle." The AI learns the shape of these particles dynamically.

3. The "Anchor" Strategy: Finding the Right Path

This is the most creative part of their discovery. When they let the AI (the neural network) try to solve the puzzle, it often gets stuck in a "fog." There are many different shapes the clay could take that almost fit the rules, but only one is the true physical reality.

• The Analogy: Imagine you are trying to find a specific house in a city where every house looks exactly the same (the "fog"). If you just walk around, you might end up at the wrong house that looks perfect.
• The Solution: The authors found that if you give the AI one single, correct piece of information about the house at a specific location (an "anchor"), the fog clears instantly.
  • Correct Anchor: If you tell the AI, "The house has a red door at this specific spot," and that is true, the AI instantly snaps into the correct solution.
  • Wrong Anchor: If you tell the AI, "The house has a blue door," the AI will still find a solution, but it will be a "fake" house that looks stable but is completely wrong.
  • The Test: The authors realized that if the solution is truly correct, the AI's answer stays very steady no matter how many times you restart the puzzle. If the anchor is wrong, the AI's answers wobble and scatter wildly. They use this "stability" to know if they have found the truth.

4. What They Tested

They tested this method on two types of puzzles:

1. Generalized Free Fields: A simplified, known type of physics theory. They used this to prove their method works. They showed that with the right "anchor," the AI could perfectly reconstruct the known answer.
2. Holographic CFTs: These are theories related to black holes and gravity (via the AdS/CFT correspondence). This is much harder. They used their method to try and find specific numbers describing these theories.
   • The Result: They found a solution that seemed stable, but when they compared it to other known methods, there was a small discrepancy (about 4% off). They admit this is likely due to the "approximate" nature of their math tools, but they proved the concept works: they can separate different types of particles (spins) that were previously impossible to untangle.

Summary

The paper introduces a new way to solve complex physics puzzles at high temperatures. Instead of ignoring the hard parts or guessing, they use mathematical shadows to handle the infinite heavy particles and AI clay models to shape the rest. Crucially, they discovered that giving the AI one correct fact (an anchor) acts like a lighthouse, guiding it out of a sea of wrong answers. If the AI's answer is steady and calm, it's likely the truth; if it's jittery, the anchor was wrong.

This is a "proceedings contribution," meaning it is a report on work in progress, sharing a new framework and early results rather than a final, perfect solution to every problem in the field.

Technical Summary

Technical Summary: Neural Networks, Dispersion Relations, and the Thermal Bootstrap

Problem Statement
The modern conformal bootstrap typically relies on positivity constraints and convex semi-definite programming to constrain Conformal Field Theories (CFTs). However, this approach fails in physically significant contexts where positivity is absent, such as finite-temperature theories, theories with defects or boundaries, and higher-point crossing equations. In these scenarios, the goal shifts from excluding regions of CFT data to reconstructing full solutions of the bootstrap equations. Standard truncations of the Operator Product Expansion (OPE) introduce systematic errors that are difficult to control, and the resulting equations often involve continuous families of constraints and an infinite number of operators, leading to non-unique solutions. Specifically, for thermal CFTs on $S^1 \times \mathbb{R}^{d-1}$, the KMS (Kubo-Martin-Schwinger) condition acts as a crossing equation that lacks positivity requirements, making the reconstruction of thermal correlators a "primal" bootstrap problem that is significantly harder than standard applications.

Methodology
The authors propose a framework that circumvents positivity and uncontrolled ad hoc truncations by combining subtracted dispersion relations with neural networks. The methodology proceeds through the following steps:

1. Dispersion Relations and High/Low-Spin Split: The thermal two-point function $g(z, \bar{z})$ is expressed via a subtracted dispersion relation. This relation separates the operator spectrum into:

   • A finite set of "exposed" low-spin, low-dimension operators ($J \le J^*$, $\Delta \le \Delta^*(J)$).
   • A set of smooth, one-dimensional "tail functions" $A_{\Delta^*(J), J}(r)$ that package the infinite tower of high-dimension contributions for each spin $J \le J^*$.
   • An integrated discontinuity term that captures all operators with spin $J > J^*$.

2. Neural Network Parametrization: The tail functions are modeled using feed-forward Multi-Layer Perceptron (MLP) neural networks. Specifically, a sparse Multi-Branch architecture is employed where a shared input layer processes the radial coordinate, feeding parallel subnets that specialize in the tail functions for different spins. This allows the infinite high-dimension spectrum to be bootstrapped dynamically alongside the exposed data.

3. Non-Convex Optimization: The KMS condition is recast as a non-convex optimization problem. The authors define loss functions (Mean Absolute Loss, Dot-Product Loss) that measure the violation of the crossing equation on a grid of points within the OPE convergence region. The optimization seeks to minimize these losses with respect to both the exposed coefficients and the neural network parameters.

4. The "Anchor" Mechanism: A critical observation is that the optimization landscape contains many low-loss minima. To select the physical solution, the framework utilizes "anchors"—specific, known values of the tail functions at an intermediate radius. The authors find that providing a correct anchor collapses the solution space to a unique, stable minimum, whereas incorrect or absent anchors lead to broad, biased distributions or failure to converge.

Key Results
The framework is tested on two distinct classes of theories:

• Generalized Free Fields (GFF): In $d=4$ with $\Delta_\phi = 1.68$, the method successfully recovers the analytic tail functions and the combination of OPE coefficients $a_{1,0} + 3a_{0,2}$ to high precision. The results demonstrate that a single correct intermediate-radius anchor is sufficient to collapse the continuous family of low-loss solutions to the unique GFF solution. Conversely, incorrect anchors result in large variances in the extracted data, establishing stability as a criterion for solution validity.
• Holographic CFTs: Applied to 4d holographic CFTs (dual to Einstein gravity in AdS), the method attempts to determine double-twist OPE coefficients from fixed multi-trace energy-momentum data. Using a ReLU-stabilized search around a reference vector (derived from GFF values), the authors extract the combination $a_{1,0} + 3a_{0,2}$. The result, $9.37 \pm 0.44$, is in tension (at the $\approx 4\sigma$ level) with independent bulk calculations ($\approx 7.7$). The authors attribute this discrepancy to systematic errors in the current implementation, specifically the truncation of the crossed-channel discontinuity and the determination of the stability tolerance parameter $p^*$.

Significance and Claims
The paper claims to provide a direct numerical bootstrap of the KMS condition at non-zero spatial separation, a regime where spin-dependent information about the thermal spectrum becomes accessible. The significance of the work lies in:

• Handling Non-Positivity: It offers a viable path for bootstrapping theories where standard positivity-based methods fail, by replacing convex optimization with non-convex optimization guided by dispersion relations.
• Dynamic Spectrum Handling: By modeling high-dimension tails with neural networks, the method avoids the systematic errors associated with hard truncations of the OPE, retaining the full infinite spectrum of contributions.
• Stability as a Criterion: The work introduces a practical strategy where the stability of the optimization (low variance across initializations) serves as a diagnostic for the correctness of the solution, provided a correct "anchor" is supplied.
• Complementarity: The authors position their approach as complementary to recent "neural spectral bias" methods (Refs. 18, 19). While those methods rely on the implicit smoothness priors of neural networks to reconstruct correlators from minimal data, this framework explicitly incorporates dispersion relations and theory-specific inputs (like the discontinuity) to resolve the full two-variable structure of thermal correlators.

The authors conclude that while the current implementation for holographic CFTs shows quantitative discrepancies with bulk results, the structural capability to disentangle spin-dependent data is confirmed. Future improvements are expected to come from increasing the spin cutoff $J^*$, implementing dynamical discontinuities, and refining the stability selection criteria.