Nonrandom network connectivity comes in pairs

Felix Z. Hoffmann, Jochen TrieschNetwork 2017 · DOI

TL;DR

This paper mathematically demonstrates that an overabundance of bidirectional connections in neural networks is a necessary consequence of nonrandom connectivity, as long as connection probabilities are symmetric within neuron pairs. This finding provides a fundamental explanation for a long-observed phenomenon in real neural circuits and suggests that bidirectional connectivity is a robust indicator of nonrandomness, rather than a specific feature to be explained by complex mechanisms.

What Problem Does This Paper Address?

Neuroscientists have long observed that the connections between neurons in the brain aren't random. Specifically, bidirectional connections (where neuron A connects to B, and B connects to A) appear far more frequently than would be expected in a purely random network. This "overrepresentation of bidirectional connections" has been noted in various cortical circuits. While this observation is well-established, the underlying principle that explains why this happens has remained elusive. The paper aims to bridge this gap by providing a theoretical framework that links nonrandomness in connection probabilities directly to the emergence of this reciprocal connectivity.

Methods

The authors employ a mathematical modeling approach. They start by defining a random graph model where each potential connection between two neurons (i and j) has a probability $P_{ij}$ of existing. Crucially, they explore two main scenarios for these probabilities:

  1. Independent Connection Probabilities: Here, the probability of connection from i to j ($P_{ij}$) is independent of the probability of connection from j to i ($P_{ji}$).
  2. Symmetric Connection Probabilities: In this scenario, $P_{ij} = P_{ji}$. This is a key assumption that reflects the idea that if distance or some other shared property influences connection likelihood, it would do so equally in both directions between a pair of neurons.

They then mathematically derive the expected probability of a bidirectional connection ($P_{bidir}$) and compare it to the expected probability in a purely random network (Erdős-Rényi graph). They use concepts like expected values, variance, and Jensen's inequality to analyze how deviations from random probabilities affect the rate of bidirectional connections. They also explore specific distributions of connection probabilities, such as a two-point distribution (where probabilities can only take two values) and a gamma distribution, to model more realistic scenarios.

Datasets

Not discussed in the paper. The study is primarily theoretical and relies on mathematical modeling and simulations.

Key Insights and Ideas

  • Symmetry is Key: The paper's central insight is that if connection probabilities are symmetric within a pair of neurons ($P_{ij} = P_{ji}$), then any nonrandomness in these probabilities necessarily leads to an overrepresentation of bidirectional connections. This is a powerful, general principle.
  • Nonrandomness Implies Reciprocity: Instead of viewing bidirectional connections as a special feature, the paper reframes them as a natural consequence of any deviation from a completely random wiring scheme, provided symmetry holds.
  • Erdős-Rényi is the Baseline: The authors use the Erdős-Rényi random graph model as a baseline. Any network that deviates from this baseline in a symmetric way will show more reciprocal connections.
  • "High" vs. "Low" Probability Pairs: Even if only a small subset of neuron pairs has a significantly higher probability of connecting than others, this can lead to a substantial overrepresentation of bidirectional connections, as long as the probabilities are symmetric within those pairs.
  • Approximate Symmetry is Enough: The overrepresentation of bidirectional connections can persist even when connection probabilities are only approximately symmetric, suggesting that real-world biological factors that introduce slight asymmetries won't negate this fundamental principle.

Key Results

  • Theoretical Proof: The paper provides a mathematical proof that for symmetric connection probabilities ($P_{ij} = P_{ji}$), a nondegenerate distribution of these probabilities (meaning not all pairs have the exact same probability) guarantees an overrepresentation of bidirectional connections ($e > 1$).
  • Quantifying Overrepresentation: Using a two-point distribution model, they show that a relative overrepresentation of $e=4$ can be achieved with a connection probability of $x=0.7$ for one group of pairs and $y=0.05$ for another, with the overall connection probability $\mu=0.1$.
  • Gamma Distribution Analysis: When modeling with a gamma distribution, they find that highly skewed distributions (where a large fraction of pairs have very low connection probabilities) can also lead to significant overrepresentation, mimicking scenarios where connectivity depends on spatial proximity.
  • Maximal Overrepresentation: For a typical cortical connection probability of $\mu=0.1$, the model predicts a maximal overrepresentation of $e=10$ under certain idealized conditions.

Limitations and Future Work

  • Idealized Models: The models used (two-point, gamma distributions) are simplifications of real neural connectivity. Future work could explore more complex, biologically realistic distributions of connection probabilities.
  • Symmetry Assumption: While the paper argues that approximate symmetry is sufficient, the core theoretical result relies on exact symmetry. Investigating the precise impact of various degrees of asymmetry on reciprocal connectivity would be valuable.
  • Higher-Order Connectivity: The paper focuses on pairwise connections. Future research could explore how these principles extend to higher-order connectivity patterns and how they interact with synaptic efficacies.
  • Inferring Computation: The authors suggest that the statistics of connectivity, like the abundance of bidirectional pairs, might offer clues about the computational principles of neural circuits. This is a promising avenue for future investigation.

Summary updated

Abstract

Overrepresentation of bidirectional connections in local cortical networks has been repeatedly reported and is a focus of the ongoing discussion of nonrandom connectivity. Here we show in a brief mathematical analysis that in a network in which connection probabilities are symmetric in pairs, <i>P</i><sub><i>ij</i></sub> = <i>P</i><sub><i>ji</i></sub> , the occurrences of bidirectional connections and nonrandom structures are inherently linked; an overabundance of reciprocally connected pairs emerges necessarily when some pairs of neurons are more likely to be connected than others. Our numerical results imply that such overrepresentation can also be sustained when connection probabilities are only approximately symmetric.
BibTeX
@inproceedings{hoffmann2017nonrandom,
  author = {Felix Z. Hoffmann and Jochen Triesch},
  title = {{Nonrandom network connectivity comes in pairs}},
  booktitle = {Network},
  year = {2017},
  doi = {10.1162/netn\_a\_00004},
}
Figure