A modeling preprint challenges a simple expectation about ecological warning signs: as a community approaches instability, pairs of species do not necessarily show near-perfect correlations in their abundances measured at the same time. In the models studied, that extreme pattern appeared only when a slowly relaxing “soft mode” — a pattern in the community that changes slowly — stood out clearly from fluctuations in the other modes.
Testing the hidden signal
To isolate the visibility effect, the researchers built a single 60-species reference community. They varied only its soft eigenvalue — the relaxation rate assigned to that mode — while holding the eigenvectors, background spectrum and noise fixed.
Visibility was defined as the per-species variance supplied by the soft mode relative to the variance from all other modes; a value of 1 marked the reported crossover. In the scan, visibility of 49 corresponded to a correlation-distribution width of 0.90, with 63% of pairs above an absolute correlation of 0.9. At visibility 0.9, the width fell to 0.26 and no pairs crossed 0.9; at 0.11, it fell to 0.07 and again none did.
The background did not disappear
The broader test used three routes for generating model communities: a synthetic feasible community, a local community assembled by immigration from a regional pool, and a community generated through repeated diversification. The main ensembles contained 600 synthetic, 250 assembled and 467 evolved communities.
Across those routes, the soft-mode signal and fluctuation background rose together. Greater spectral depth therefore did not yield an isolated dominant soft direction. Median visibility was 0.84 [0.62, 1.15] in synthetic communities, 1.44 [0.95, 2.34] in assembled communities and 1.28 [0.96, 1.82] in evolved communities; the ranges are interquartile ranges.
A stronger signal, but no universal marker
Organization raised the strongest pairwise links without making them universal. The median highest absolute pairwise correlation was 0.383 in synthetic communities, 0.515 in assembled communities and 0.508 in evolved communities. Size-matched synthetic controls had medians of 0.374 at 32 species and 0.221 at 80 species.
The researchers also compared the measured strongest pair with a prediction based on variance alone. The measured value was 0.46 of that prediction in synthetic communities, 0.65 in assembled communities and 0.48 in evolved communities. The remaining modes contributed mainly negative terms: median mode-diagonal contributions were −0.61, −0.48 and −0.60, while cross terms were small and positive at 0.04, 0.06 and 0.06.
Near-perfect pairs were possible but rare. Four assembled communities, or 1.6% of that ensemble, had a strongest pair with an absolute correlation above 0.9. None of the 600 synthetic communities, 467 evolved communities or 900 control communities did.
The model’s boundaries
Those findings come from models rather than field observations. The core setup used a symmetric interaction matrix with unit self-interaction and environmental noise divided into matched and idiosyncratic components, while the dynamics were analyzed in a linearized form. The result therefore does not establish a universal rule for real ecosystems or for other ecological models.
The document is arXiv preprint version 1, dated 20 August 2026. Within the studied models, its central warning is limited but important: a strong same-time correlation is not, by itself, a universal measure of how close a community is to instability.
Paper data and sources
Original title: Correlations at criticality in ecological communities
Authors: Akiva Goldberg, Nadav M. Shnerb
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text