Notes on algebraic perspective on neural network conformal theories

In the vast landscape of modern physics, there is a persistent effort to understand how the fundamental rules of the universe emerge from the behavior of countless tiny components. One powerful way physicists study these rules is through a framework called conformal field theory, which describes systems that look the same regardless of how much you zoom in or out. These theories are usually built on the idea of a "Hilbert space," a mathematical stage where every possible state of a system has a positive, measurable size, much like how a physical object has a positive mass. This positivity is crucial because it ensures the theory makes sense in the real world, preventing probabilities from becoming negative or nonsensical. However, nature sometimes presents us with systems that break these standard rules, leading to what are known as non-unitary theories. For decades, these systems were difficult to study because the standard mathematical tools used to build them would collapse without that guarantee of positive size.

Recently, a new bridge has been built between this abstract world of physics and the practical world of artificial intelligence. Researchers have discovered that if you take a very large collection of simple neural networks—computer models inspired by the human brain—and average their behavior, the resulting patterns mimic the behavior of these exotic physical theories. This connection, known as the neural network-field theory correspondence, suggests that the statistical properties of the neural network's internal settings can generate the exact same mathematical structures as complex quantum fields. This offers a fresh, non-traditional way to study physics, one that does not rely on the usual equations of motion but instead on the statistics of the network's parameters.

A team of researchers has now taken a closer look at this connection, specifically focusing on a class of these neural network theories that produce non-unitary conformal field theories. These are the tricky systems where the standard rules of positivity fail, and the usual mathematical stage for the theory seems to disappear. The researchers wanted to understand the underlying structure of these theories: how the different parts of the system relate to one another, and whether a consistent mathematical framework could still be built even when the standard rules of positivity are broken. They were particularly interested in a puzzle that had emerged in earlier work: while the theory produced consistent patterns for how particles interact, it seemed to contain an infinite number of identical, or "degenerate," states that were hard to distinguish. It was unclear if this was a flaw in the model or a genuine feature of the physics.

To solve this, the team examined the mathematical "Gram matrix," a tool that measures how much different parts of the system overlap with one another. In a standard theory, this matrix must be positive definite, meaning it always yields positive numbers, which guarantees that the states are distinct and well-behaved. The researchers found that for their neural network theories, this matrix is indeed positive definite when measured in the space of the neural network's parameters. This means that within the context of the neural network itself, every operator is distinct and well-defined. However, when they looked at the theory from the perspective of space and time, the positivity vanished. The states that were distinct in the parameter space became mixed and negative in the space-time description, confirming that these are indeed non-unitary theories that cannot be described by a standard Hilbert space.

The team then constructed a new algebraic framework to describe these theories. Instead of relying on the standard Hilbert space, which requires positive sizes for all states, they built a structure based on a "Krein space." This is a more flexible mathematical stage that allows for both positive and negative sizes, provided they are balanced in a specific way. They showed that even without the strict positivity of a standard quantum theory, the correlations between different points in space and time still obey the symmetries of conformal field theory. They demonstrated that the infinite tower of degenerate operators found in earlier studies is likely not an artificial glitch but a natural consequence of how the neural network parameters average out, a conclusion supported by numerical trends and conjectured for the infinite limit. The researchers showed that these operators are related to one another in a precise way, forming a coherent algebraic structure that can be described using the symmetries of the conformal group, though they noted that proving the full Krein positivity of this structure remains an assumption rather than a derived fact.

One of the most significant findings was the resolution of a puzzle regarding "mixed" interactions. In previous studies, it was unclear how to calculate the interaction between different types of operators when they were degenerate. The team showed that the standard method of simply multiplying operators together does not work here. Instead, the correct way to understand these interactions is to view them as an expansion around a set of "root" operators, where the connections between different states are mediated by the specific statistical distribution of the neural network parameters. They found that the constraints on the statistical moments of the neural network parameters are deeply linked to the consistency of the theory. If the distribution of parameters is chosen correctly, the theory remains consistent, even though it lacks the standard positivity of a physical Hilbert space.

The researchers also clarified the distinction between the "algebra of fields" and the "algebra of correlators." In the neural network construction, the fields themselves, before averaging, follow a simple multiplication rule. However, once the averaging is performed to get the physical predictions, the resulting correlations reveal a much richer and more complex structure. This structure is defined by the conformal symmetry and the specific way the neural network parameters are distributed. The team showed that this algebraic structure is robust and can be defined without needing to assume the existence of a standard Hilbert space. Instead, they used the state-operator correspondence, a fundamental principle in conformal field theory, to define the algebra directly from the correlation functions.

In their conclusion, the authors emphasized that while these theories do not fit into the standard framework of unitary quantum field theory, they are not mathematically broken. They exist as valid algebraic structures that can be studied and understood. The work provides a recipe for constructing the algebra of these theories directly from their correlation functions, bypassing the need for a traditional Hilbert space. This opens the door to studying a wider class of physical systems that were previously considered too difficult to handle. The researchers noted that while they have established the existence of this algebraic structure, the full details of how these systems behave in a local, space-time sense are still being explored. They suggested that the large-N limit, where the number of neural networks becomes very large, provides a clean setting to test these ideas further.

Ultimately, this work demonstrates that the tools of artificial intelligence can be used to construct and analyze complex physical theories that lie outside the standard boundaries of physics. By treating the neural network parameters as the fundamental ingredients, the researchers were able to build a consistent mathematical description of non-unitary conformal field theories. They showed that the apparent degeneracy and lack of positivity are not errors but features of a new kind of physical theory that can be described using a Krein space. This finding suggests that the landscape of possible physical theories is broader than previously thought, and that the intersection of machine learning and theoretical physics offers a powerful new way to explore the fundamental laws of nature. The study confirms that even in the absence of standard positivity, a coherent and structured theory can emerge from the statistics of neural networks, providing a new perspective on how physical laws might be encoded in complex systems.