Lattice Field Theory for a network of real neurons

Abstract: In a paper [Bardella et al., Entropy 26 (6), 495 (2024)] we introduced a simplified Lattice Field Theory (LFT) framework that allows experimental recordings from major Brain-Computer Interfaces (BCIs) to be interpreted in a simple and physically grounded way. From a neuroscience point of view, our method modifies the Maximum Entropy model for neural networks so that also the time evolution of the system is taken into account and it can be interpreted as another version of the Free Energy principle (FEP). This framework is naturally tailored to interpret recordings from chronic multi-site BCIs, especially spike rasters from measurements of single neuron activity.

• The paper introduces a novel lattice field theory framework that extends MaxEnt models to incorporate non-stationary dynamics in real neural networks.
• It employs a discrete QFT formulation to infer synchronous and temporal couplings from high-dimensional BCI data while maintaining causality.
• The method unifies statistical mechanics, Bayesian inference, and free energy principles to yield a scalable model for complex neural recordings.

Lattice Field Theory Applied to Networks of Real Neurons

Introduction and Motivation

This work presents a minimalistic Lattice Field Theory (LFT) framework that systematically extends the Maximum Entropy (MaxEnt) approach to empirical neural data by incorporating explicit non-stationary time evolution. The traditional MaxEnt models, as previously employed in neural data analysis, are isomorphic to the equilibrium Ising model and thus fundamentally constrained to stationary statistics [Schneidman et al., Nature 2006]. Experimental protocols using chronic multi-site Brain-Computer Interfaces (BCIs)—notably, spike rasters from multi-electrode arrays—produce time-resolved, high-dimensional ensemble data. Existing field-theoretic formulations appear either too abstract for direct application to such observables or demand physicist-level technical expertise, impeding wider use in neuroscience. The proposed Neural LFT eliminates dependence on high-energy physics formalisms by providing a concrete, computationally tractable discrete-time, binary-field formalism tailored for real-world BCIs. Importantly, it synthesizes connections with thermodynamic and Bayesian principles, yielding tractable inference procedures for interaction parameters.

Extension Beyond Stationarity: Quantum Field Theory Embedding

Classical MaxEnt frameworks interpret neural population activity through a stationary correlation structure, effectively reducing inference to fitting pairwise (Ising-like) interactions at a fixed time window. This formalism precludes meaningful treatment of neural dynamics or memory effects. Drawing on the Parisi-Wu stochastic quantization method, the proposed LFT treats the neuronal system as a set of binary fields ($\Omega_i^\alpha \in \{0,1\}$ indexed by neuron $i$ and time $\alpha$) defined on a discrete spacetime lattice. The action functional $\mathscr{A}(\Omega)$ defines a Gibbs measure on the $2^{NT}$-dimensional configuration space—akin to a Euclidean QFT partition function—where the physical time parameter is preserved, and a "fictional" time direction is introduced for technical quantization.

This construction allows the explicit modeling of non-stationary, temporally causal interactions. The statistical mechanics (SM) framework is recovered as a ground-state, stationary limit ($\hslash \to 0$), while the quantum regime naturally incorporates dissipative and temporal effects. Thus, the LFT directly addresses the critical challenge articulated in recent reviews: constructing tractable models that link time-extended neural observables to underlying couplings [Meshulam & Bialek, RMP 2025].

Neural Lattice Action and Inference Structure

A neural system of $N$ units over $T$ time points is represented as a binary bit string, mapping the full ensemble to a lattice field. The core of the LFT is the action expansion,

$$\mathscr{A}(\Omega|\,F,I) = \sum_{i,\alpha} I^\alpha_i \Omega^\alpha_i + \sum_{i,j,\alpha,\beta} F^{\alpha\beta}_{ij} \Omega^\alpha_i \Omega^\beta_j + \cdots$$

where $F$ tensors encode interaction strengths, with higher-order vertex terms truncated for computational tractability and empirical justification (based on the smallness of observed higher-order correlations). The two-vertex truncation yields a model strictly interpretable in terms of space (synchronous) and time (causal, memory) couplings. Parameter inference is cast in terms of matching "grand covariance" matrices—encompassing both spatial and temporal covariances—which modern inversion methods can efficiently handle [Berg et al., Adv. Phys. 2017].

The model enforces causality by requiring $i$0 for $i$1, and "local memory" by further constraining off-diagonal blocks: only within-neuron (self) interactions extend across time, while cross-neuron interactions are purely synchronous. Implementing "bi-stationarity" (stationarity in either neuron or time indices), the parameter space is dramatically reduced from $i$2 to $i$3, with further halving due to symmetry.

The ultimate form of the action reflects three types of terms:

• External field (input) terms $i$4,
• Synchronous (space-like) couplings $i$5,
• Temporal (time-like, memory) couplings $i$6.

These map one-to-one to the means and covariances of the empirical neural activity kernel.

Free Energy Principle and Bayesian Link

A central theoretical contribution is the explicit connection to the Free Energy Principle (FEP), operationalized via the variational bound on model evidence. Jensen's inequality on the partition sum shows that the action-induced free energy functional $i$7 bounds the negative log partition function, leading directly to a variational Bayesian interpretation of inference and to the unification of FEP-compliant agent theory with QFT models. The neural LFT thus provides a principled route to model selection and parameter estimation within the Bayesian Brain paradigm [Friston, Neural Computation 2017].

Application to High-Dimensional Experimental Data

The framework is validated on data from Utah 96 arrays, which record the spiking activity of up to 100 neurons organized spatially over a cortical lattice [Pani et al., PNAS 2022]. The mapping of measured data to the observable kernel $i$8 proceeds via spatial decimation, carefully matching the discretization scale (columnar organization) to neuroanatomical features. The inference of $i$9 and $\alpha$0 is both numerically feasible and interpretable. Empirically, the degree distribution and scaling properties of $\alpha$1 can be aligned with anatomical constraints (e.g., sparse vs. dense, connectivity exponents $\alpha$2 between 0 and 1). Notably, the time couplings $\alpha$3 capture memory and refractoriness, with their structure reflecting the dynamical refractory period, including its adaptation to varying population activity levels.

Theoretical and Practical Implications

The Neural LFT framework:

• Unifies several established models (Ising/MaxEnt, PCA, Hopfield) as limiting cases.
• Provides tractable inference of both instantaneous and memory couplings from realistic BCI data.
• Demonstrates a rigorous, physically-justified approach to model selection and regularization via the free energy functional.
• Offers a scalable path for modeling high-dimensional, temporally-resolved neuroscience datasets that escape stationary assumptions.

Limitations include the need for further validation on complex long-duration recordings where higher-order memory effects are nontrivial, and the computational burden in cases where $\alpha$4 and $\alpha$5 become very large.

Perspectives and Future Directions

Potential developments include analyzing more complex BCI datasets with longer, non-stationary recordings to reveal nontrivial temporal covariance structures and genuine long-term memory effects. The application of the LFT to layers of artificial neural networks, enabling a direct comparison between biological and deep learning systems, is an intriguing prospect—especially in the context of developing interpretable representations of deep activations (the "kernel of activations" or hypermatrix structures).

Benchmarking and further analytical validation may exploit exactly solvable long-memory binary dynamics, such as the Elephant Random Walk or HLS processes. Cross-disciplinary exchange—bringing the sophisticated renormalization tools of particle physics to bear on computational neuroscience—promises future methodological advances.

Conclusion

The presented Neural Lattice Field Theory delivers a robust, theoretically sound, and computationally accessible extension of MaxEnt approaches, embedding empirical neural data analysis within a discrete QFT framework that naturally incorporates non-stationary dynamics and memory effects. The empirical mapping to real BCI recordings and compatibility with the Bayesian FEP paradigm broaden its utility, both for theoretical investigations and practical data analysis in neuroscience and AI. The reduction of parameterization, clear causal structure, and compatibility with high-dimensional, high-throughput datasets highlight its relevance for the future development of computational and theoretical neuroscience.