Bayes or Heisenberg: Who(se) Rules?

Volker Tresp, Hang Li, Federico Harjes, Yunpu MaarXiv 2025 · arXiv · OpenReview

Abstract

Although quantum systems are generally described by quantum state vectors, we show that in certain cases their measurement processes can be reformulated as probabilistic equations expressed in terms of probabilistic state vectors. These probabilistic representations can, in turn, be approximated by the neural network dynamics of the Tensor Brain (TB) model. The Tensor Brain is a recently proposed framework for modeling perception and memory in the brain, providing a biologically inspired mechanism for efficiently integrating generated symbolic representations into reasoning processes.

Introduction

Two major theories currently dominate the landscape of cognitive science: the Bayesian Brain hypothesis, which posits the brain as a statistical inference engine, and Quantum Cognition, which utilizes the mathematical framework of quantum mechanics to explain human decision-making patterns that appear "irrational" under classical logic.

However, both frameworks face significant hurdles. Exact Bayesian inference is often computationally intractable for biological systems, requiring power and time beyond what a brain can provide. Conversely, the "Quantum Brain" hypothesis is frequently dismissed by physicists because the biological brain is too "warm and wet" to maintain the delicate quantum coherence required by standard quantum mechanics. This paper introduces a "probabilistic quantum" framework that bridges this gap, reformulating quantum measurements as neural network dynamics within the Tensor Brain (TB) model.

The Tensor Brain Framework

The Tensor Brain model proposes that the brain can achieve the mathematical elegance of quantum mechanics using classical, probabilistic neural hardware. This is achieved through several key innovations:

  • Probabilistic Quantum States: Instead of complex-valued state vectors, the model utilizes probabilistic state vectors. This bypasses the need for physical coherence while retaining the structural advantages of quantum logic.
  • Pro-bits: These are the probabilistic analogues of qubits, implemented as formal neurons (Bernoulli variables) that represent the system's state.
  • Unitary-Stochastic Matrices: To ensure the system remains within valid probability bounds during state evolution, the authors employ a specific class of matrices that combine unitary properties with stochastic constraints.
  • Neural Approximations: To maintain scalability, the model replaces high-dimensional tensor contractions with efficient vector-matrix multiplications using Jensen’s approximation and standard neural network layers.

The HB-POVM Protocol

A central contribution of this research is the HB-POVM (Heisenberg-Bayes Positive Operator-Valued Measure). In standard quantum mechanics, a Projective Valued Measure (PVM) causes a state to "collapse," often discarding previous information. In contrast, the HB-POVM preserves state information, acting effectively like a Bayesian update.

This protocol allows the system to integrate new evidence without losing the context of its prior state. The authors demonstrate that in the absence of "postselection" (filtering for specific outcomes), the HB-POVM update is mathematically equivalent to a Bayesian update in a generative Hidden Markov Model (gHMM), yet it remains computationally tractable where the gHMM does not.

Insights into Modern AI Architectures

The TB framework provides a novel theoretical lens for understanding common components in artificial intelligence:

  1. Skip Connections as Logit Priors: The additive "skip connections" found in ResNets and Transformers can be interpreted as the mathematical incorporation of a "logit prior." This represents the brain’s mechanism for preserving its previous state while integrating new sensory evidence.
  2. Attention as Ignorant Measurement: The attention mechanism in AI is reframed as a form of "ignorant" quantum measurement—a process where the outcome is not explicitly recorded but the measurement itself alters the system's state to resolve contextual ambiguity.

Experimental Results

The model was evaluated on its sensitivity to "order effects"—the phenomenon where the sequence of information changes the final conclusion—using a subset of ImageNet (100,000 samples) and WordNet hierarchies.

  • Stability: The HB-POVM model demonstrated significantly higher stability than standard PVM. When the order of queries was reversed, the Kullback-Leibler (KL) divergence for HB-POVM was only 0.304, compared to a highly volatile 20.54 for PVM.
  • Consistency: Qualitatively, the HB-POVM model maintained semantic consistency. For instance, when identifying a "Chihuahua," the model's conclusion remained stable regardless of whether it was first asked about the broader "Animal" category. In contrast, PVM models often suffered from state collapse, leading to erroneous classifications like "Tarantula" depending on the query sequence.

Conclusion and Future Work

This work argues for an "ontic" interpretation of the brain: probabilistic sampling is not merely a result of missing information, but an intrinsic part of the brain's fundamental hardware. While the model is biologically inspired, mapping these tensor operations to specific dendritic or synaptic processes remains a challenge. Future research will focus on the "h-layer" bottleneck—the compression of information into a dynamic context—and the potential for applying the HB-POVM framework to Large Language Models to improve reasoning and reduce hallucinations.

Summary updated

Abstract

Although quantum systems are generally described by quantum state vectors, we show that in certain cases their measurement processes can be reformulated as probabilistic equations expressed in terms of probabilistic state vectors. These probabilistic representations can, in turn, be approximated by the neural network dynamics of the Tensor Brain (TB) model. The Tensor Brain is a recently proposed framework for modeling perception and memory in the brain, providing a biologically inspired mechanism for efficiently integrating generated symbolic representations into reasoning processes.
BibTeX
@misc{tresp2025bayes,
  author = {Volker Tresp and Hang Li and Federico Harjes and Yunpu Ma},
  title = {{Bayes or Heisenberg: Who(se) Rules?}},
  year = {2025},
  eprint = {2510.13894},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.13894},
}
Figure