Bayes or Heisenberg: Who(se) Rules?

Basic Information

• Paper ID: 2510.13894
• Title: Bayes or Heisenberg: Who(se) Rules?
• Authors: Volker Tresp, Hang Li, Federico Harjes, Yunpu Ma
• Institution: LMU Munich
• Classification: q-bio.NC (Neurons and Cognition), cs.AI, cs.LG, quant-ph
• Publication Date: October 17, 2025
• Paper Link: https://arxiv.org/abs/2510.13894

Abstract

This paper demonstrates that the measurement process of quantum systems can be reformulated as probabilistic equations based on probabilistic state vectors under specific conditions. These probabilistic representations can be approximated through the neural network dynamics of the Tensor Brain (TB) model. Tensor Brain is a framework for modeling brain perception and memory, providing a biologically-inspired mechanism for efficiently integrating generated symbolic representations into reasoning processes.

Research Background and Motivation

Core Questions

This research attempts to answer fundamental questions: How does information processing in the brain work? How are perception and memory generated? Should entirely new AI modeling and reasoning approaches be considered?

Research Motivation

1. Interdisciplinary Integration Requirement: Combining quantum theory, Bayesian reasoning, and neural networks to provide a unified framework for brain information processing
2. Computational Complexity Issues: Traditional Bayesian reasoning has prohibitively high computational complexity in practical applications, necessitating more efficient alternatives
3. Biological Plausibility: Existing Bayesian brain hypotheses face criticism for biological implausibility, requiring models more consistent with biological mechanisms

Limitations of Existing Approaches

• Bayesian Brain Hypothesis: While theoretically elegant, exact Bayesian inference is computationally intractable and typically requires variational approximations
• Quantum Brain Theory: Faces strong criticism because quantum coherence times in the brain are extremely short, insufficient to support meaningful quantum computation
• Pure Neural Network Approaches: Lack theoretical foundations and struggle to explain symbolic reasoning and memory mechanisms

Core Contributions

1. Probabilistic Quantum Framework: Proposes a method to reformulate quantum measurement processes as probabilistic equations, replacing quantum states with probabilistic states
2. Heisenberg-Bayes POVM (HB-POVM): Introduces a novel measurement scheme that preserves state information and establishes connections between quantum and Bayesian measurements
3. Pro-bits Concept: Proposes probabilistic analogues of qubits, implementing probabilistic computation through unified random matrices
4. Neural Network Implementation: Demonstrates that probabilistic quantum algorithms can be efficiently implemented through neural networks, forming the basis of the Tensor Brain algorithm
5. Equivalence Theorem: Proves that under specific conditions, probabilistic HB-POVM updates are mathematically equivalent to Bayesian updates in generative Hidden Markov Models (gHMM)
6. Computational Advantages: Under post-selection conditions, probabilistic HB-POVM inference remains tractable while gHMM inference becomes intractable

Methodology Details

Task Definition

The core task is to establish bridges between quantum theory, probabilistic reasoning, and neural computation, providing a unified mathematical framework for brain perception and memory mechanisms.

Theoretical Framework

1. Quantum Operations Foundations

• Quantum State Evolution: $\psi \leftarrow U_{evol}\psi$, where $U_{evol}$ is a unitary evolution operator
• Projective-Valued Measure (PVM): Measurement probability $P(k) = |u_k^\dagger\psi|^2$, posterior state $\psi \leftarrow u_k$
• Positive-Operator-Valued Measure (POVM): More general measurement description allowing noise and environmental interaction

2. HB-POVM Innovation

Defining diagonal measurement operators: $M_{k}^{HB} = \text{diag}(u_k)$

Measurement probability: $P(k) = b_k^T p$

Posterior update: $\psi \leftarrow \frac{u_k \circ \psi}{\sqrt{P(k)}}$

where $\circ$ denotes Hadamard product and $b_k$ is the $k$-th column of unitary random matrix $B$.

3. Probabilistic Quantum Representation

Through an "ignorance measurement" protocol, replacing quantum states with probabilistic states:

• Probabilistic State Evolution: $p \leftarrow B_{evol}p$
• Probabilistic PVM: $p \leftarrow b_k$
• Probabilistic HB-POVM: $p \leftarrow \frac{b_k \circ p}{b_k^T p}$

4. Pro-bits and Neural Implementation

Tensorizing probabilistic states into pro-bits (probabilistic bits): $p_{i(i_1,...,i_n)} = \prod_{\ell=1}^n \gamma_\ell^{i_\ell}(1-\gamma_\ell)^{1-i_\ell}$

where $\gamma_\ell = \text{sig}(q_\ell)$, implemented through neural network approximation.

Tensor Brain Algorithm

Core Algorithm

1. Evolution Algorithm: $h \leftarrow \text{sig}(v_0 + Vq)$, $q \leftarrow Wh$
2. Input and Attention: $q \leftarrow q + g(\nu) + \sum_k a_k \text{softmax}(\cdot)$
3. Generative Measurement: Sample $k$, update $q \leftarrow \alpha q + \beta a_k$

Key Characteristics

• Cognitive Brain State (CBS): $\gamma = \text{sig}(q)$ represents activation of the representation layer
• Symbolic Index Layer: Encodes symbolic interpretations, interacting with CBS to produce perception and memory
• Skip Connection Interpretation: HB-POVM update $q \leftarrow a_k + q$ provides theoretical explanation for skip connections

Experimental Setup

Datasets

• ImageNet Subset: 100,000 samples from 200 fine-grained and 16 coarse-grained categories
• WordNet Hierarchy: Establishes hierarchical relationships between fine-grained and coarse-grained labels

Evaluation Metrics

• KL Divergence: Measures differences in prediction distributions across different query orders
• Jensen-Shannon Divergence (JSD): Symmetrized version of KL divergence
• Label Reversal Rate: Proportion of label changes across different query orders

Comparison Methods

• PVM Model: Standard projective-valued measure model
• HB-POVM Model: Proposed Heisenberg-Bayes POVM model

Implementation Details

• ResNet-50 for visual feature extraction
• Category embeddings computed as averages of extracted feature vectors
• Each image annotated with regions of interest (ROI)

Experimental Results

Main Results

Query Order Effect Experiment

Key Findings

1. Order Invariance: HB-POVM maintains high consistency across different query orders, while PVM exhibits strong order dependence
2. Quantum Interference: Quantum PVM demonstrates interference effects; probabilistic PVM eliminates interference while preserving order effects
3. Computational Advantages: Under post-selection conditions, HB-POVM inference remains tractable while classical Bayesian updates become intractable

Case Analysis

Qualitative analysis on Chihuahua images shows:

• PVM: Query order affects predictions; fine-grained→coarse-grained yields (Chihuahua, animal), coarse-grained→fine-grained yields (tarantula, animal)
• HB-POVM: Both query orders stably predict (Chihuahua, animal)

Related Work

Quantum Cognition

• Uses mathematical structures of quantum theory to describe human reasoning biases
• Non-commutative algebra captures order effects and violations of the law of total probability

Bayesian Brain Hypothesis

• Assumes the brain implements Bayesian reasoning at multiple cognitive levels
• Faces criticism regarding computational intractability and biological implausibility

Quantum Computing and the Brain

• Penrose-Hameroff theory proposes quantum computation in microtubules
• Faces strong criticism due to extremely short quantum coherence times in the brain

Conclusions and Discussion

Main Conclusions

1. Unified Framework: Successfully establishes a unified framework connecting quantum theory, probabilistic reasoning, and neural computation
2. Biological Plausibility: Provides information processing mechanisms more consistent with biology than pure Bayesian approaches
3. Computational Advantages: Achieves computational tractability while maintaining theoretical rigor
4. Broad Applicability: Framework applies to multiple cognitive functions including perception, memory, and symbolic reasoning

Limitations

1. Approximation Assumptions: Neural network implementation requires multiple approximation assumptions (conditional independence, Jensen approximation, etc.)
2. Limited Experimental Validation: Validation primarily focuses on order effects, lacking verification across broader cognitive tasks
3. Biological Correspondence: Correspondence between theoretical framework and actual brain mechanisms requires further verification

Future Directions

1. Extended Experiments: Validate framework effectiveness across more cognitive tasks
2. Neuroscience Verification: Seek correspondences between framework predictions and neuroscience experimental results
3. Engineering Applications: Apply framework to practical AI system design

In-Depth Evaluation

Strengths

1. Theoretical Innovation: First to establish direct connections between quantum measurement and neural computation, proposing original concepts like HB-POVM
2. Mathematical Rigor: Provides complete mathematical derivations and proves key equivalence theorems
3. Interdisciplinary Integration: Successfully synthesizes concepts and methods from physics, neuroscience, and computer science
4. Practical Value: Offers new approaches to addressing computational complexity in Bayesian reasoning
5. Biological Inspiration: Provides a framework more consistent with biological principles than traditional approaches

Shortcomings

1. High Complexity: Theoretical framework is complex, requiring deep background in quantum mechanics and neuroscience for complete understanding
2. Insufficient Experimental Validation: Experiments concentrate on order effects, lacking comprehensive verification of other framework aspects
3. Approximation Errors: Multiple approximation steps may accumulate errors, affecting final performance
4. Interpretability: While providing theoretical foundations, practical interpretability in applications remains to be improved

Impact

1. Academic Value: Provides new theoretical perspectives for cognitive science and AI, potentially catalyzing new research directions
2. Practical Prospects: Offers theoretical guidance for designing more efficient reasoning algorithms
3. Interdisciplinary Impact: May promote further cross-disciplinary fusion between quantum information, neuroscience, and AI

Applicable Scenarios

1. Symbolic Reasoning Systems: Suitable for AI systems requiring interaction between symbolic and sub-symbolic information
2. Memory Modeling: Applicable to semantic and episodic memory modeling
3. Perception Systems: Suitable for systems requiring integration of multimodal perceptual information
4. Large Language Models: Can provide theoretical foundations and improvement directions for transformer architectures

References

This paper cites important literature from multiple domains including quantum information theory, cognitive science, and neural networks, including Nielsen & Chuang's quantum computing textbook, Busemeyer & Bruza's quantum cognition monograph, and original Tensor Brain papers.

Overall Assessment: This is an important interdisciplinary research paper with significant theoretical value, successfully unifying quantum theory, probabilistic reasoning, and neural computation within a single framework. Despite theoretical complexity and limited experimental validation, its originality and potential impact make it an important contribution to cognitive science and AI.