Descending into the Modular Bootstrap

This paper employs machine-learning-style optimization, featuring a novel singular-value-based optimizer and uncertainty estimation, to numerically explore the landscape of two-dimensional conformal field theories with central charges between 1 and 8/7, identifying candidate spectra in a previously uncharted region and suggesting a tighter constraint on the spectral gap near $c=1$.

The Gist

Imagine the universe is built from a giant, infinite library of blueprints. These blueprints describe how matter and energy behave at the most fundamental level. In physics, these blueprints are called Quantum Field Theories (QFTs).

For a long time, physicists have been trying to catalog every single valid blueprint in this library. But here's the problem: most of these blueprints are incredibly complex, and we don't have a complete list of what's allowed and what's forbidden.

This paper is about a team of physicists who decided to stop waiting for a librarian to hand them a list. Instead, they built a robotic explorer (using machine learning) to wander through the library, looking for new, valid blueprints that no one has ever seen before.

Here is the story of their journey, broken down into simple concepts.

1. The Goal: Finding "Valid" Blueprints

In the world of 2D physics (a simplified version of our universe), there are strict rules a blueprint must follow to be real. The most important rule is called Modular Invariance.

Think of this rule like a perfectly balanced mobile hanging from the ceiling. If you twist the mobile (change the perspective), it must look exactly the same. If the blueprint doesn't balance perfectly when twisted, it's a fake blueprint—it doesn't describe a real universe.

The physicists wanted to find new blueprints that balance perfectly. They focused on a specific, mysterious region of the library where the "central charge" (a number that measures the complexity of the blueprint) is between 1 and 1.14. It's a "no-man's-land" where we know no blueprints exist yet, but we suspect some might be hiding there.

2. The Problem: The Library is Too Big

The library is infinite. You can't check every single page. So, the physicists had to make a shortcut: they decided to only look at the first few chapters of the blueprint (the most important, low-energy parts) and ignore the rest.

This is called Truncation.

The Analogy: Imagine trying to judge a whole novel by only reading the first 50 pages.

• The Risk: You might think the story is great, but maybe the ending is terrible.
• The Fix: The authors invented a new way to estimate how much they are "missing" by only reading the first 50 pages. They created a "Uncertainty Score." If their guess is shaky, the score goes up. If their guess is solid, the score stays low.

3. The Tool: The "Sven" Robot

To find these hidden blueprints, they needed a smart robot. Standard robots (called "Gradient Descent") are like hikers who only look at the ground directly in front of them. If they are in a deep valley with steep walls, they get stuck and can't find the deepest, best spot.

The authors used a new robot called Sven.

• The Analogy: Imagine you are in a foggy, mountainous valley. A normal hiker walks straight down the steepest slope. Sven, however, has X-ray vision. It can see the entire shape of the valley at once. It knows that to get to the bottom, it sometimes needs to take a step sideways or even slightly uphill to navigate a tricky ridge.
• The Result: Sven is much better at finding the deepest, most stable valleys (the best blueprints) than the standard hikers.

4. The Discovery: A "Gap" in the Library

The team sent Sven to explore the "no-man's-land" (Central Charge 1 to 1.14).

What they found:

1. Success: They found several new, valid blueprints! These aren't just random numbers; they look and behave exactly like the known blueprints of the universe. This suggests that the "no-man's-land" isn't empty; it's actually full of a continuous stream of valid theories.
2. The Mystery Gap: However, they found a strange hole. In a specific corner of the library (where the complexity is low but the "spectral gap" is high), Sven couldn't find any valid blueprints, even though the old rules said they should be allowed there.

The Analogy: Imagine you are looking for a specific type of tree in a forest. You find them everywhere, except in one small, sunny clearing. The old map says trees should grow there. But your robot says, "I've looked everywhere, and there are no trees here."

• The Conclusion: This suggests the old map is wrong. There is a hidden rule (related to the fact that particles must come in whole numbers, not fractions) that prevents trees from growing in that specific clearing.

5. Why This Matters

This paper is a big deal for three reasons:

1. New Physics: It proves that there are likely many more types of universes (theories) than we thought, filling in the gaps between the ones we already know.
2. Better Tools: They showed that using Machine Learning with "Uncertainty Scores" and the "Sven" robot is a powerful new way to solve hard physics problems. It's like upgrading from a magnifying glass to a super-computer.
3. The "Integer" Rule: They discovered that the requirement for particles to come in whole numbers (integers) is much stricter than we thought. It acts like a filter, blocking out certain theories that we previously thought were possible.

Summary

The authors built a smart, X-ray-vision robot to search a vast library of theoretical universes. They found that the library is much fuller than we thought, but there is a specific "forbidden zone" where the rules of whole numbers prevent any universe from existing. This discovery helps us understand the fundamental laws of nature a little better, proving that even in the chaotic quantum world, there are strict, hidden patterns.

Technical Summary

1. Problem Statement

The paper addresses the challenge of mapping the landscape of two-dimensional conformal field theories (2d CFTs), specifically focusing on the "irrational" regime where theories possess an infinite number of primary operators and no conserved currents beyond the stress tensor.

• The Gap: While unitary 2d CFTs are exhaustively cataloged for central charge $c=1$ and $c<1$, there is no known exhaustive catalog for $c > 1$. Specifically, the range $c \in (1, 8/7)$ is devoid of confirmed, exactly solvable examples.
• The Primal Approach: The authors adopt a "primal" approach to the modular bootstrap. Instead of deriving bounds on operator dimensions (the "dual" approach using semi-definite programming), they attempt to explicitly construct candidate partition functions that satisfy modular invariance.
• Key Challenges:
  1. Truncation: The modular bootstrap involves an infinite number of constraints. Numerical solutions require truncating the spectrum at a maximum dimension $\Delta_{\text{max}}$, introducing significant uncertainty.
  2. Loss Landscape: The space of solutions is complex, with hierarchical scales that cause standard optimization algorithms (like gradient descent) to get stuck in local minima.
  3. Integer Constraints: Physical CFTs require integer-valued degeneracies for operators, a constraint that is difficult to enforce in continuous optimization.

2. Methodology

The authors translate the modular bootstrap equations into a machine-learning-style optimization problem.

A. Formulation as Optimization

• Objective: Minimize the violation of modular invariance for the torus partition function $Z(\tau) = Z(-1/\tau)$.
• Parameters: The central charge $c$ and the spectral gap $\Delta_{\text{gap}}$ are fixed. The optimization variables are the conformal dimensions $\{\Delta_a\}$ of a set of primary operators with fixed spins and degeneracies (initially set to 1).
• Loss Function: Instead of a standard Mean Squared Error (MSE), the authors introduce a $\chi^2$-style loss function:
  $$L = \frac{1}{|T|} \sum_{\tau \in T} \left( \frac{Z(\tau) - Z(\tau^S)}{\sigma_{\text{trunc}}(\tau)} \right)^2$$
  Here, $\sigma_{\text{trunc}}(\tau)$ is a theoretically derived uncertainty estimate for the truncation error. This normalization ensures that a loss value of $O(1)$ corresponds to "CFT-like" behavior, regardless of the truncation scale $\Delta_{\text{max}}$.

B. Technical Innovations

1. Truncation Uncertainty Estimation:
   • The authors propose a deformation strategy to estimate $\sigma_{\text{trunc}}$. They take a known baseline CFT (e.g., a free boson at $c_0=1$) and deform its partition function to mimic the asymptotic density of states of a CFT with $c > c_0$.
   • By truncating this deformed partition function at $\Delta_{\text{max}}$ and comparing it to the untruncated version, they derive a $\tau$-dependent uncertainty estimate. This allows the loss function to be robust against the choice of $\Delta_{\text{max}}$.
2. Sven Optimizer:
   • Standard gradient descent fails to navigate the hierarchical loss landscape (where scales vary exponentially due to Virasoro characters).
   • The authors utilize the Sven algorithm, a singular-value-based optimizer. It performs a Singular Value Decomposition (SVD) of the Jacobian matrix of residuals.
   • By keeping only the top $k$ singular values (where $k \approx 16$), Sven can traverse multiple parameter directions simultaneously, effectively navigating "canyons" in the loss landscape that gradient descent misses.

C. Validation

The methodology is validated using known $N=1$ Minimal Models ($m=5$ to $9$). The authors demonstrate that:

• The improved loss function yields values of $O(1)$ for known CFTs.
• The individual loss terms follow a $\chi^2$ distribution.
• The parameter uncertainties (derived from the inverse Hessian of the loss) reveal a continuous space of solutions with highly correlated parameters.

3. Key Results

A. Candidate CFTs in the "Gap"

The authors successfully identified candidate partition functions in the previously empty region $c \in (1, 8/7)$.

• They found high-quality solutions for $c \in \{1.05, 1.07, 1.09, 1.11, 1.13\}$.
• These candidates exhibit spectra visually similar to known Minimal Models, with clusters of operators and spectral gaps.
• Uncertainty analysis suggests these candidates are part of a continuous space of solutions, implying that modular invariance alone does not force discreteness in this region.

B. The Primal/Dual Gap

A comprehensive survey of the $(c, \Delta_{\text{gap}})$ plane revealed a significant obstruction:

• Observation: While the dual modular bootstrap bound (HCLY bound) allows for $\Delta_{\text{gap}} \leq c/6 + 1/3$, the authors' primal optimization failed to find any solutions in the region $c \in (1.00, 1.06)$ and $\Delta_{\text{gap}} \in (0.3, 0.5)$.
• Significance: This "primal/dual gap" suggests a stricter constraint on the spectral gap than currently known from the dual approach.

C. Role of Integrality and Truncation

The authors investigated the origin of this gap:

• Integrality: When they relaxed the constraint of integer-valued degeneracies (allowing continuous degeneracies), the gap closed completely, and solutions were found up to the HCLY bound. This indicates the gap is driven by the requirement that operator degeneracies must be integers.
• Truncation Scale ($\Delta_{\text{max}}$):
  • At $\Delta_{\text{max}} = 3$, the gap is small or non-existent.
  • At $\Delta_{\text{max}} = 4$, the gap opens significantly.
  • At $\Delta_{\text{max}} = 5$, the gap persists and the excluded region expands.
  • This suggests that the integrality constraints on higher-spin operators ($j \geq 2$) are the primary drivers of the new, tighter bounds on the spectral gap.

4. Significance and Conclusions

• New Constraints: The paper provides strong numerical evidence for a more stringent bound on the spectral gap near $c=1$ than the existing HCLY bound, driven specifically by the integrality of operator degeneracies.
• Methodological Advancement: The work demonstrates the efficacy of ML-style optimization (specifically the Sven algorithm) combined with principled uncertainty estimation for solving complex, hierarchical physics problems.
• Abundance vs. Rarity: The existence of a continuous space of solutions in the $c \in (1, 8/7)$ range suggests that consistent small-$c$ CFTs are abundant, at least when considering only modular invariance.
• Future Directions: The authors suggest that the "primal/dual gap" is a real physical feature that requires further investigation via the dual bootstrap (imposing integrality on higher-spin operators) and potentially exploring crossing symmetry constraints to fully characterize these new theories.

In summary, this paper successfully uses a novel optimization framework to construct candidate CFTs in a previously unexplored regime, revealing that the requirement for integer degeneracies imposes tighter constraints on the spectral gap than previously understood.