Source
Saeedian, Pigani, Maritan, Suweis & Azaele (2021), LIT-057, the arXiv preprint, read in full in NOTE-061. The published version in Phil. Trans. R. Soc. A 380(2227) was not read, and its abstract is reworded.
What was actually shown
The setting. Generalised Lotka–Volterra dynamics with delayed interactions only (A = 0, Eq. 12). Off-diagonal entries are Normal(0, σ_B) with connectance C, the diagonal is −d_B, and growth rates are U[0.05, 0.1]. Every delay figure uses d_B = 10, σ_B = 0.1, C = 1. Runs start in the coexistence region, where the undelayed equilibrium is feasible and globally stable (Appendix A).
The critical delay (Eq. 14, Appendix C). Each community eigenvalue j gives λ = j e^{−λτ}. The equilibrium loses stability at τ_c, the minimum over the spectrum, by a Hopf bifurcation. At S = 100 the community is stable at τ = 12, oscillates at τ = 18 and diverges numerically at τ = 28 (Figs. 2–4). τ_c and the divergence threshold both fall with S (Fig. 5b). Below τ_c, resilience falls as τ grows.
Variability (Eq. 15, Fig. 6). At τ = 20 and 22, the coefficient of variation of total abundance, CV_c, falls with S over about S = 10–70, while the species-averaged CV_s rises. In the stable phase both are zero by construction.
What could have come out otherwise. The simulated onset could have missed the analytic τ_c. It matched, at the one setting.
What this does not say
- That interaction complexity sets τ_c. In this strongly self-regulated regime the community eigenvalues are about −r_i, so τ_c ≈ π/(2 max r_i), about 16 for r ≤ 0.1, which brackets the observed onset between 12 and 18. The fall with S may then be the largest growth rate rising with the number drawn. That is NOTE-061’s inference, unverified, and the paper calls the trend consistent with May.
- That diversity stabilises the community. Summing S weakly correlated oscillators lowers CV_c roughly as 1/√S for statistical reasons. The paper cites Tilman et al. (1998) on exactly this null and does not test against it. Realisations and error bars for Fig. 6 are not stated.
- That the oscillations stay bounded. The divergence past τ_c^C persisted when the Euler step was refined. The Conclusion mentions “analytical insights” on boundedness that the paper does not contain, and the Introduction says it does not prove boundedness.
- Anything with both instantaneous and delayed interactions. No result combines them.
- The arctan branch. Eq. 14 leaves it unstated; for a nearly real eigenvalue it must be read so as to give the classical π/(2|α|).
Connections
- Non-reciprocal metabolic cycles (THEORY-634) is the other case in this group where oscillation comes from the structure of an interaction matrix: a cycle’s parity removes the real mode, as a delay here rotates a stable pair across the axis.
- The BZ reaction (THEORY-508) oscillates by a relaxation cycle in a chemical medium. Here the oscillation needs no autocatalysis, only a lag in the response.
Note of 2026-10-20: the derivation, and a re-run against the null
The critical delay. With every interaction delayed, the bracket of Eq. 12 vanishes at x* = −B⁻¹r, so δx′(t) = J δx(t − τ) with J = diag(x*)B, and on each eigenvector λ = j e^{−λτ}. A root λ = iω needs ω = |j| and i e^{iωτ} = j/|j|, so τ_j = (arg j − π/2)/|j| = arctan(|Re j|/|Im j|)/|j| for Re j < 0, which is π/(2|j|) for real j and fixes the branch Eq. 14 leaves open. At τ = 0 all roots are the j; new roots of a retarded equation enter from Re λ = −∞; so stability is first lost at τ_c = min_j τ_j. From λ − j e^{−λτ} = 0, dλ/dτ = −λ²/(1 + λτ), and at λ = iω its real part is ω²/(1 + ω²τ²) > 0, so the pair crosses transversally. Supercriticality is not derived.
The re-run. The paper’s setting (d_B = 10, σ_B = 0.1, C = 1, r ~ U[0.05, 0.1], feasible draws only), Euler with dt = 0.05 to t = 4000, the first 2000 discarded, S = 10, 20, …, 70, τ = 20 and 22, 20 realisations per point. The null shifts each species’ series by an independent random circular lag and sums them.
| S | CV_c (τ=20) | CV_c/null | CV_s | CV_c (τ=22) | CV_c/null | CV_s |
|---|---|---|---|---|---|---|
| 10 | 0.174 | 0.98 | 0.39 | 0.243 | 1.02 | 0.55 |
| 30 | 0.097 | 0.82 | 0.46 | 0.136 | 0.94 | 0.63 |
| 50 | 0.082 | 0.89 | 0.48 | 0.109 | 0.93 | 0.65 |
| 70 | 0.072 | 0.87 | 0.52 | 0.103 | 0.98 | 0.70 |
Every ratio is within one standard deviation (about 0.2) of 1, and CV_c√S stays near 0.55–0.60. Onset matched τ_c (S = 50: flat at τ = 12, CV_c 4·10⁻⁴ at 15 against τ_c,min = 15.1, 0.055 at 18), and dt = 0.02 moved CV_c by 3%. One aside: τ_c is not the growth-rate bound π/(2 max r) NOTE-061 suggested; from S = 40 its correlation with that bound is about 0.2, and it falls below it (13.1 against 15.8 at S = 200), so at larger S the interactions set it.