Non-random network connectivity comes in pairs

Felix Z. Hoffmann, Jochen TriescharXiv 2016 · DOI · arXiv

Abstract

Overrepresentation of bidirectional connections in local cortical networks has been repeatedly reported and is in the focus of the ongoing discussion of non-random connectivity. Here we show in a brief mathematical analysis that in a network in which connection probabilities are symmetric in pairs, $P_{ij} = P_{ji}$, the occurrence of bidirectional connections and non-random structures are inherently linked; an overabundance of reciprocally connected pairs emerges necessarily when the network structure deviates from a random network in any form.

Introduction

Neuroscientists have long observed that local cortical circuits are highly structured, deviating significantly from the uniform randomness of Erdős-Rényi graphs. One of the most prominent "non-random" features is reciprocity: the observation that if neuron A connects to neuron B, it is much more likely that B also connects to A than chance would suggest. While researchers have used this overrepresentation as a primary indicator of non-randomness, the exact mathematical relationship between the two hasn't been fully clear. This paper demonstrates that if connection probabilities are symmetric ($P_{ij} = P_{ji}$), then any deviation from a uniform random network must result in an overabundance of bidirectional pairs.

Theoretical Framework and Methods

The authors utilize a random graph model where each potential connection between two nodes ($i$ and $j$) has a specific probability $P_{ij}$ of existing. A key assumption is symmetry in connection probabilities, where the likelihood of a connection from $i \to j$ is identical to that from $j \to i$. This assumption is biologically grounded: factors like spatial distance or functional similarity (e.g., orientation tuning) are inherently mutual.

Symmetric connection probability model illustrating the two-point distribution where pairs have either high ($x$) or low ($y$) connection probabilities.

To analyze the impact of connectivity "shape" on reciprocity, the study examines two probability distributions:

  1. Two-point distribution: A simplified "toy model" where pairs are either "highly likely" or "unlikely" to connect.
  2. Gamma distribution: A continuous distribution capable of modeling the skewed connectivity often observed in biological networks.

The core mathematical insight relies on Jensen’s Inequality. The probability of a bidirectional connection is the expected value of the square of the connection probability, $E[P_{ij}^2]$. Since $x^2$ is a strictly convex function, Jensen’s Inequality dictates that $E[P_{ij}^2] \ge (E[P_{ij}])^2$. Consequently, any variance in connection probabilities across the network—a hallmark of non-random structure—necessarily leads to higher reciprocity than a uniform random network.

Key Results

The analysis yields several critical findings regarding network structure:

  • Universal Lower Bound: For any non-degenerate symmetric distribution, the relative occurrence of bidirectional pairs ($\rho$) is always greater than 1.
  • Maximal Overrepresentation: The theoretical maximum for $\rho$ is $1/\mu$, where $\mu$ is the overall connection probability. In cortical networks with $\mu \approx 0.1$, the maximum possible overrepresentation is 10x.
  • Skewness and Cortical Data: Using the Gamma distribution, the authors demonstrate that highly skewed distributions—where a small fraction of pairs have high connection probabilities—readily reproduce the ~4x overrepresentation observed in rat visual cortex data.
Analysis of the Gamma distribution model showing how skewness (lower shape parameter $\alpha$) leads to greater overrepresentation $\rho$.

Discussion and Limitations

These results suggest that bidirectional overrepresentation should be viewed not as a standalone "goal" of neural wiring, but as an inevitable "symptom" of underlying organizational principles like spatial clustering or functional hubs.

While the math explains why these pairs appear, it doesn't explain their functional role. Furthermore, the study focuses strictly on pairs of neurons and does not address higher-order motifs, such as three-neuron loops or cliques. Future work may extend this analysis to higher-order structures and dynamic connectivity changes driven by synaptic plasticity.

Summary updated

Abstract

Overrepresentation of bidirectional connections in local cortical networks has been repeatedly reported and is in the focus of the ongoing discussion of non-random connectivity. Here we show in a brief mathematical analysis that in a network in which connection probabilities are symmetric in pairs, $P_{ij} = P_{ji}$, the occurrence of bidirectional connections and non-random structures are inherently linked; an overabundance of reciprocally connected pairs emerges necessarily when the network structure deviates from a random network in any form.
BibTeX
@misc{hoffmann2016non,
  author = {Felix Z. Hoffmann and Jochen Triesch},
  title = {{Non-random network connectivity comes in pairs}},
  year = {2016},
  eprint = {1609.04245},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1609.04245},
}
Figure