Contribution
A self-contained cavity calculation for the random generalized Lotka–Volterra model of a single community fed by migration from a species pool, without assuming symmetric interactions. It gives the phase diagram (unique fixed point, multiple attractors, unbounded growth), the diversity and the abundance distribution in closed form in the first phase, and a change of variables that shows the fixed points depend on the parameters only through three combinations, which explains why the unique-fixed-point boundary is independent of the mean interaction strength.
Key insight
Once abundances are normalized by their mean, the fixed-point problem depends only on u ≈ 1/σ, the symmetry γ and the scaled spread of carrying capacities σ_λ. Adding one more species to a large community perturbs the others only by linear response, so the new species sees a Gaussian field and its abundance is that field divided by a renormalized self-interaction u − γv, truncated at zero. The unique fixed point stops being stable when its response to small perturbations diverges; past that line the community has more than one possible outcome, and which it reaches depends on its assembly history.
Assumptions
- Dynamics (1): dN_i/dt = (r_i/K_i) N_i (K_i − N_i − Σ_{j≠i} α_ij N_j), noise neglected.
- α_ij = μ/S + σ a_ij with ⟨a_ij⟩ = 0, ⟨a_ij²⟩ = 1/S, ⟨a_ij a_ji⟩ = γ/S; otherwise independent, any distribution. K_i independent of α, mean 1, Gaussian with variance σ_K² when not identical.
- Large S with μ, σ, γ fixed (“individually weak interactions”). The calculation assumes an uninvadable fixed point exists and checks the assumption self-consistently; it uses feasibility and uninvadability, not dynamical stability, which the author checks numerically (Mij* = α_ij r_i N_i / K_i positive definite in the UFP phase for several distributions of r).
- Migration acts on time scales long enough for the community to relax between invasions.
Key results
- Symmetry (§III). Fixed points satisfy 0 = n_i(λ_i − u n_i − Σ a_ij n_j + h), with n_i normalized abundances, u = (1 − μ/S)/σ, λ_i = (K_i − 1)/(σ⟨N⟩) and h = (1/⟨N⟩ − μ)/σ. For equal K_i the normalized properties do not depend on μ. These fixed points are those of a replicator equation, though the full dynamics differ.
- Abundance distribution (9). n = max(0, (h + √(q + σ_λ²) z)/(u − γv)), z standard normal: a truncated Gaussian. Self-consistency (10)–(11) fixes q = ⟨n²⟩, v, h and the surviving fraction φ. ⟨N⟩ = 1/(σh + μ), and unbounded growth begins where σh + μ = 0.
- Loss of stability (13). ⟨(δn)²⟩/⟨ξ²⟩ = φ/((u − γv)² − φ), which diverges at φ = (u − γv)², equivalently u = (1 + γ)v. For σ_K = 0 this is σ_c = √2/(1 + γ) for all μ > 0 (Appendix A). Lower symmetry and more variable carrying capacities both move the transition to larger σ.
- Simulations. S = 15, 100, 200, 400, 800; the predictions for φ and the moments hold at S = 200 in the UFP phase and approximately into the MA phase; at S = 15 with communities of about six species the fit is still good. The fraction of systems with more than one final community jumps at σ_c for large S and is broadened at finite S, with unique fixed points found “well inside” the MA phase even at S = 100.
- MA phase. For γ = 0 the dynamics may reach non-fixed-point attractors and, with an extinction cutoff, may never become uninvadable; invasions can switch the community between alternatives.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | In the UFP phase the large-S abundance distribution is a Gaussian truncated at zero | strong | cavity derivation §IV; simulations Figs. 3–4 |
| C2 | The UFP–MA transition is at φ = (u − γv)², i.e. σ_c = √2/(1 + γ) for equal carrying capacities, independent of μ | strong | (13), Appendix A; re-derived; Fig. 5(b) |
| C3 | Past the transition the community outcome depends on assembly history | moderate | numerical: repeated runs from different initial conditions, Fig. 5(b); not derived |
| C4 | In the UFP phase every feasible uninvadable fixed point is also dynamically stable | moderate | numerical check over several r distributions, §V |
| C5 | Feasibility, uninvadability and stability all fail together at one transition | moderate | §VI, interpretive summary of C2–C4 |
Method
The cavity method: add species 0 with fresh interactions; treat its effect on the others by linear response with susceptibilities v_ij = ∂n_i/∂ξ_j; the diagonal sum gives the self-interaction correction γv and the off-diagonal terms vanish at large S. The stability calculation perturbs every λ_i by independent noise and asks when the variance of the response diverges.
Concepts
- UFP phase: a unique, stable, uninvadable fixed point for every realization.
- MA phase: multiple dynamical attractors, fixed or not; composition depends on initial conditions or invasion order.
- Persistent species: N_i > 0 at the fixed point; φ is their fraction of the pool.
Connections
The equations match those of the random replicator model of Diederich and Opper and Opper and Diederich (refs. 4, 6) for σ_λ = 0, and the transition is of the same kind as theirs. Rieger (ref. 5) treated the Lotka–Volterra case under assumptions that hide the MA phase. The companion arXiv paper (Bunin 2016, 1607.04734) uses the same model for the structure of assembled interaction networks. Blumenthal, Rocks and Mehta (LIT-1236) apply the same cavity stability test to a consumer–resource model with non-reciprocal interactions and identify its transition with this paper’s MA transition. CLAIM-369 is the symmetric case of a related model, where dynamics are gradient flow.
Bearing on the record
- QUESTION-083. The LIT files this as the phase in which outcomes depend on history. The reading supports that, with a qualification: what is derived is where the unique outcome ends; that the alternative outcomes then depend on order of arrival is shown by simulation only, and for γ = 0 the dynamics may not settle at all. “Entrenchment of whatever gained influence early” (the synthesis’s reading) is a reading of the MA phase that the paper does not test: it does not study which of several attractors early arrivals select.
- CLAIM-369. For γ = 1 (symmetric interactions) this model has MacArthur’s Lyapunov function V = Σ K_i N_i − ½ Σ α_ij N_i N_j (a standard fact for positive r_i/K_i, checked here, not stated in the paper), and the MA phase still exists (σ_c = 1/√2). So symmetry and an optimization principle do not by themselves give a unique outcome; multiplicity of minima does the work. This is consistent with CLAIM-369, which concerns the A = UᵀU case, positive semidefinite and so convex.
No instruction for machine-learning practice.
Limitations
- The MA phase is located but not analysed; the author says it “may require more powerful and involved theoretical tools”.
- No noise, no spatial structure, a single trophic level, unstructured random interactions. The truncated-Gaussian abundances differ from the lognormal or power-law forms commonly observed, which the author attributes to omitted processes.
- Errors and slips. Typographical only: “posses”, “at at large S”, “is not longer consistent”, “resistent”. None in the derivations checked.
Open questions
- In the MA phase, how many attractors are there and how do their basins depend on invasion order? The paper points to later work.