Quantum Mechanics and Neural Networks

The Big Picture

Imagine tracking a dust particle suspended in water: a jittering, unpredictable zigzag too jagged for any smooth description. Now imagine that buried inside that erratic motion is the complete mathematical structure of quantum mechanics, with built-in uncertainty, discrete energy levels, and the rule that certain properties can’t all be precisely pinned down at the same time. That’s the claim at the heart of new work from Northeastern University’s Christian Ferko and James Halverson. Any quantum mechanical system can be represented as a neural network, and the weirdness of quantum theory emerges directly from the roughness of the network’s paths.

Quantum mechanics and neural networks have been converging for years, with researchers using neural nets to solve quantum equations, accelerate simulations, and approximate quantum states. Ferko and Halverson flip the question. Instead of asking how AI can help us study quantum systems, they ask: are quantum systems already neural networks? The answer, they prove, is yes, in a mathematically precise sense.

They show that every valid quantum mechanical theory, formulated in imaginary time (a standard technique that makes quantum calculations more tractable), can be written as a collection of neural networks with a specific probability distribution over their parameters. This isn’t an approximation or a metaphor. It’s a theorem.

Any quantum mechanical theory in Euclidean time is mathematically equivalent to a neural network. The full structure of quantum mechanics, unitarity and uncertainty included, can be encoded in network architectures and parameter distributions.

How It Works

The bridge runs through stochastic processes, mathematical descriptions of randomly evolving quantities like that dust particle’s position over time. In Euclidean quantum mechanics, a particle’s trajectory through imaginary time looks like a random walk. This random walk satisfies the conditions for the Kosambi-Karhunen-Loève (KKL) theorem, a result from probability theory that decomposes a random process into a sum of simpler components.

That expansion is exactly a neural network. The network’s neurons correspond to basis functions in the KKL decomposition, its weights to random coefficients, and the distribution over those weights defines the ensemble. Three conditions seal the deal:

• Mean-square path continuity: quantum paths don’t jump discontinuously
• Finite two-point functions: quantum correlation functions are well-behaved
• The Kállén-Lehmann spectral representation: the energy spectrum is non-negative

These properties hold for every legitimate quantum mechanical system. So every quantum system already lives inside a neural network.

Correlation functions computed in imaginary time must also satisfy reflection positivity to guarantee a valid real-time quantum theory. Reflection positivity is the imaginary-time expression of unitarity, the requirement that probabilities add up to one. So how do you make sure a neural network has this property?

Ferko and Halverson identify two routes. A parameter splitting method divides the network’s parameter space into two halves, one for positive times and one for negative, and constrains their relationship so that reflection positivity is baked in. The second route is more elegant: Markov processes (random walks where only the current state matters) are automatically reflection positive. This turns out to be the more productive approach.

Deep Neural Network Quantum Mechanics

Take any Markov process and feed it through a neural network. The output loses the Markov property (the network introduces memory, mixing past and present) but stays reflection positive. You can stack neural networks on top of Markov processes and always remain within the space of valid quantum theories.

The authors call this deep neural network quantum mechanics (deep NN-QM): constructing families of quantum systems by composing neural networks on top of a base quantum process. Each layer transforms the quantum system into a new one while preserving unitarity.

They test this numerically using the Ornstein-Uhlenbeck process, a particle pulled back toward equilibrium by a restoring force (think of a mass on a spring, but randomized), as their seed. Applying various architectures and computing observables directly from the resulting paths, they confirm:

• Heisenberg uncertainty: position-momentum uncertainty saturates the quantum lower bound
• Non-trivial commutators: $[x, p] = i\hbar$ arises because paths are nowhere differentiable, too rough to have well-defined slopes. This roughness is how non-commutativity enters the theory.
• Energy spectrum: reconstructed eigenvalues recover the correct harmonic oscillator spectrum

Passing the Ornstein-Uhlenbeck process through increasingly complex architectures (deeper networks, different activation functions) produces genuinely different quantum systems. The machinery works.

Why It Matters

This result ties together two active programs in modern physics: the neural network field theory program, which interprets machine learning through the lens of quantum field theory, and constructive quantum mechanics, which asks what minimal conditions define a valid quantum theory. These aren’t separate projects. They are two faces of the same structure.

The payoff runs in both directions. If every quantum system has a neural network representation, the full toolkit of machine learning (optimization algorithms, architecture search, expressivity theory) becomes available for studying quantum systems without ever writing down an action or Lagrangian. Going the other way, insights from quantum theory about which network ensembles are physically meaningful could inform the design of better neural architectures.

Open directions include extending the framework to quantum field theories in higher dimensions, mapping specific architectures to classes of quantum systems, and understanding how training dynamics modify quantum properties.

For decades, physicists have used neural networks as tools to probe quantum systems. This paper argues the relationship is deeper than that: quantum mechanics and neural networks are, in a precise mathematical sense, the same kind of object.

Ferko and Halverson prove that every quantum mechanical theory is a neural network, and every deep neural network built on a quantum seed defines a new quantum system, unified through a theorem about random paths and their roughness.

---