Grid cells compute distance via de-correlation of heterogeneous populations

A team led by Pritipriya Dasbehera, Akshunna S. Dogra, and William T. Redman from arXiv (q-bio.NC, 2025/2026) has developed a mathematical framework for how grid cells—neurons in the medial entorhinal cortex—compute distance via de-correlation. Inspired by evidence that grid cell populations exhibit small but robust heterogeneity in their spatial firing properties, the researchers hypothesized that distance can be derived simply from the decorrelation of population activity as an animal moves. They built a 1D theory predicting that, even with noisy grid cells, distance encoding emerges naturally. Simulations confirmed the theory, and surprisingly revealed a non-intuitive 'sweet spot': distances that are farther apart can be better encoded than some nearer distances. They found preliminary evidence for this pattern in previously published rodent behavioral experiments, and showed that a decoder estimating distance from de-correlation produced a similar error profile. This challenges the intuition that nearby locations should always be easier to distinguish.

Extending the theory to two dimensions, the team simulated noisy grid cells and uncovered a fundamental trade-off. As heterogeneity in grid properties increases, the range of distances that can be encoded expands, but the distinguishability between different distances decreases. Conversely, less heterogeneity improves distinguishability over a shorter range. Remarkably, the average amount of grid property variability measured in real neural populations strikes an optimal balance, enabling encoding of distances up to several meters. This work provides new mechanistic insight into spatial navigation and explains why grid cells are slightly heterogeneous—it's not noise, but a feature that maximizes distance coding efficiency. The findings could also inform AI models for spatial reasoning and path integration.