Distance by de-correlation: Computing distance with heterogeneous grid cells

The Big Picture

Close your eyes and imagine walking from your front door to the corner store. You don’t consciously count your steps, yet your brain somehow knows, roughly, how far you’ve traveled. Now imagine doing this without sight, smell, or landmarks. Just the raw sense of distance, computed on the fly. Neuroscientists have been puzzling over how this works for decades.

The leading candidate for the brain’s internal odometer is a class of neurons called grid cells, found in a region called the medial entorhinal cortex (MEC), a navigation hub near the hippocampus. These neurons fire in a repeating hexagonal pattern as an animal moves through space. Think of a sheet of graph paper overlaid on the environment, where each intersection is a firing field. The prevailing theory held that encoding distances beyond a meter or so required the brain to combine signals from multiple grid cell clusters (modules), each tuned to a different scale. A kind of neural arithmetic where reading the combination tells you how far you’ve gone. It worked in theory, but it assumed an unrealistically tidy neural architecture, and nobody could show how downstream brain regions would actually decode the signal.

Pritipriya Dasbehera, Akshunna Dogra, and William Redman now propose something far simpler: the brain doesn’t need that multi-module machinery at all. A single module of grid cells, with just a small amount of natural variability in their firing patterns, can encode distance through plain statistical de-correlation. Neural activity patterns gradually become less similar to each other the farther apart two locations are.

Key Insight: When grid cells within a single module have small but real differences in their grid spacing, the similarity between neural activity patterns decreases monotonically with distance, giving the brain a readable odometer that requires no complex cross-module decoding.

How It Works

The core idea rests on a mathematical property of populations of oscillating neurons. Start with a perfectly uniform grid module: every cell fires with the exact same spacing and orientation, just offset in phase. When you compare the population vector (the collective pattern of activity across all cells) at two locations separated by exactly one grid spacing, the correlation snaps back to 1. Correlation here just measures how similar two activity patterns are, on a scale from 0 (nothing in common) to 1 (identical).

It’s déjà vu, built into the math. The neural code has no memory of how far you’ve come; it only knows where you are within the repeating grid.

Now add a small dose of heterogeneity: let each cell’s grid spacing vary slightly around a common mean, as recent experiments confirmed actually happens in real brains. The code changes fundamentally:

• The quasi-periodic “snapping back” breaks down
• Population vector correlations no longer return to 1 at integer multiples of the grid spacing
• Instead, correlations decay monotonically with distance: smaller correlation reliably means greater distance
• A single grid module, with modest variability, now acts as a true odometer

The team developed a formal mathematical theory for the one-dimensional case, describing how correlations behave as a function of distance and variability in grid spacing. Simulations of noisy synthetic grid cells matched the theoretical predictions cleanly.

But the simulations turned up something stranger. You might expect a simple rule: the farther you go, the harder it is to encode distance precisely. The model instead predicts a non-monotonic pattern. Encoding fidelity decreases, then increases, then decreases again as distance grows. There is a mid-range sweet spot where some longer distances are actually encoded more reliably than shorter ones just before them.

This sounds like a bug. It turns out it might be a feature already hiding in the literature. The team found exactly this pattern in published rodent behavioral experiments: mice performed better on certain intermediate distances than on shorter ones. When the researchers ran their decoder on simulated grid cell populations, it produced a closely matching error profile. The framework didn’t just explain existing data; it predicted a specific, counterintuitive pattern that was already sitting in the numbers, waiting to be noticed.

Real navigation is two-dimensional, of course. Extending to 2D simulations, the same basic principle holds, but with a wrinkle. There is a trade-off between two things you’d want from a distance code:

• Range: how far a distance you can encode before the correlation signal becomes too flat to read
• Distinguishability: how well the code separates nearby distances from each other

The amount of variability in grid properties controls where you land on this trade-off curve. Too little, and there’s no usable de-correlation. Too much, and the code loses structure; nearby distances become hard to tell apart.

The average observed heterogeneity in real rodent grid cells (roughly 10–20% variability in grid spacing) lands almost exactly at the balance point. It’s a Goldilocks result: enough variability to produce a monotonic distance signal, not so much that the code dissolves into noise. This puts the encoding range at several meters, precisely the right scale for a rat navigating a real environment.

Why It Matters

Why does biological neural architecture look the way it does? Random variation in neural properties is typically treated as noise to be tolerated, not a feature to be explained. Dasbehera, Dogra, and Redman’s framework flips that assumption. The heterogeneity of grid cells isn’t a biological imperfection. It may be the specific ingredient that makes distance coding possible without multi-module integration.

Current GPS-inspired theories of spatial cognition, which often invoke elaborate hierarchical multi-scale codes, may be solving a harder problem than the brain actually faces. A simple de-correlation principle, applied to a single heterogeneous population, achieves the same goal with fewer moving parts. That’s worth considering for anyone designing spatial representations in artificial systems, whether navigation algorithms or memory architectures in neural networks.

The work raises sharp empirical questions too. Can the de-correlation signal be directly measured in MEC recordings during active navigation? Do grid cells with more variability cluster in regions associated with longer-range navigation? And how does this mechanism interact with the place cells downstream that actually guide behavior?

Bottom Line: A single grid module with naturally observed levels of variability encodes distance through simple statistical de-correlation, no complex multi-module machinery required. The resulting sweet spot in encoding fidelity matches real rodent behavior, offering a cleaner and more biologically plausible theory of spatial navigation.