Contribution
It takes the consumer–resource model whose symmetric form has a minimization principle and breaks the symmetry between what a species gains from a resource and what it does to it. A cavity calculation gives the steady state and locates where it stops being stable, and that boundary turns out to depend on one ratio: how many species survive per resource left undepleted. Numerics show the other side is chaotic.
Key insight
With e = c the effective species interaction matrix Aᵢⱼ = Σ_α cᵢα eⱼα Θ(R_α) is symmetric and the dynamics descend an objective, so they reach a unique fixed point. Make c and e only partly correlated and A is not symmetric; no objective is guaranteed and fixed points can lose stability. Where they do is set by species packing: the more species the undepleted resources must support, the more reciprocity is needed to hold the fixed point, ρ* = √(S*/M*).
Assumptions
- Dynamics (1)–(2): dNᵢ/dt = Nᵢ(Σ_α cᵢα R_α − mᵢ), dR_α/dt = R_α(K_α − R_α) − Σᵢ Nᵢ eᵢα R_α. Self-replenishing (logistic) resources.
- Thermodynamic limit S, M → ∞ with M/S fixed; cᵢα = μ_c/M + σ_c dᵢα/√M, eᵢα = μ_e/M + σ_e(ρdᵢα + √(1 − ρ²) xᵢα)/√M, so corr(cᵢα, eⱼβ) = ρ δᵢⱼ δ_αβ; K and m with independent fluctuations. Figures use μ_c = μ_e, σ_c = σ_e.
- Replica-symmetric, self-averaging cavity ansatz. The paper does not prove it; it is checked against simulations for moderate non-reciprocity (Figs. A6, A8, not read).
- Simulations add immigration λ_N = λ_R = 10⁻¹⁰ to regularize and to keep chaos from being killed by finite-size extinctions.
Key results
- Phase boundary, eq. (6)/(B18). The fixed point is unstable when ρ < ρ*, ρ*² = (number of surviving species)/(number of non-depleted resources), evaluated at the boundary. It is a co-dimension-one surface in parameter space, since those numbers are set by the parameters.
- ρ = 1. Competitive exclusion keeps the ratio ≤ 1, so the instability is never reached; the instability and the cavity’s infeasibility transition coincide there.
- Infeasibility. For some parameters the cavity equations have no solution (χ → 0, where the surviving species equal the non-depleted resources); this lies inside the unstable phase and is not physical. At ρ = 0 the only cavity solution is total extinction.
- Chaos (§Chaos, App. C). With S = M = 512, largest Lyapunov exponents are positive in runs that do not settle and negative in runs that do; |λ₁| ≈ 5 × 10⁻³ for the Fig. 2 parameters. GALI decays exponentially in the dynamic phase. Kaplan–Yorke dimension 15 ± 7 against 74 ± 17 surviving species and resources, close to the 38 ± 19 that switch between high and low abundance.
- Finite size (App. E). Fitting times-to-steady-state with a censored maximum-likelihood model at two parameter points, the estimated probability of ever reaching steady state stays ≈ 1 in the stable phase and falls towards 0 with size in the dynamic phase (fit ξ = 278.5); the time scale grows as M^0.4 and M
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | The replica-symmetric fixed point of the asymmetric MCRM loses stability at ρ² = S/M* | strong, given the ansatz | App. B; re-derived from the cavity relations |
| C2 | Below ρ* the dynamics are chaotic | moderate | numerical: Lyapunov exponents, GALI, two trajectories; at one parameter point mostly |
| C3 | In the thermodynamic limit no system in the dynamic phase reaches steady state | moderate | finite-size MLE fits at two parameter points, S, M ≤ 1024 |
| C4 | The symmetric MCRM never reaches this transition | strong | competitive exclusion bound on S*/M* ≤ 1 with ρ* ≤ 1 |
| C5 | The divergence of c from e generalizes Tilman’s two-species stability condition | weak | analogy in the discussion |
Method
Cavity: add one species and one resource, treat the perturbation of the rest by linear response with susceptibilities χ (resource to carrying capacity) and ν (species to mortality), and close self-consistent equations for survival fractions and moments (A39–A47). Stability: perturb all survivors by independent noise and compute the second moments of the responses; their divergence marks the boundary. Numerically the equations are solved by black-box least squares.
Concepts
- Asymmetric MCRM (aMCRM): the model above, with impacts eᵢα distinct from benefits cᵢα.
- Reciprocity ρ: the correlation between a species’ benefit from and impact on each resource.
- Species-packing ratio: surviving species over non-depleted resources.
- Dynamic phase: the side of the boundary with persistent fluctuations.
Connections
It applies to a consumer–resource model the stability test Bunin used for random Lotka–Volterra communities (LIT-1234), and the SI identifies its transition with that paper’s multiple-attractor transition. The symmetric model’s minimization principle is MacArthur’s, cited through Marsland, Cui and Mehta’s minimum environmental perturbation principle (the record’s LIT-1230) and Chesson (1990). In the record, CLAIM-369 proves the symmetric, resource-mediated case to be gradient flow on a convex objective; this paper is where that picture’s assumption fails.
Bearing on the record
- QUESTION-083 and the synthesis. The LIT’s reason for being here, that “with non-reciprocal interactions the dynamics need not minimize any objective”, is what the paper argues, and the reading supports it with a sharper condition: loss of reciprocity alone does not produce chaos; it does so only when ρ falls below √(S*/M*). A community well below the competitive exclusion bound tolerates more asymmetry before it goes unstable. So “the limit of reading distributed agency as optimization” is not a point but a boundary that depends on packing.
- CLAIM-369. The claim’s gradient-flow form needs A = UᵀU. This paper gives the case A = CEᵀ with C ≠ E and shows what fails, for a related consumer–resource model; it is the natural objection thread to the claim’s scope, though not to the claim, which states its symmetric assumption.
- THEORY-1162 is drawn from this reading.
No instruction for machine-learning practice.
Limitations
- The boundary rests on the replica-symmetric ansatz; beyond it the ansatz is invalid and the dynamic phase is characterized by simulation only, at a few parameter points and moderate sizes.
- Logistic, self-replenishing resources; one trophic level of consumers; random unstructured interactions.
- Errors and slips. (i) The fast-resource reduction in the letter writes R_α = max{0, K_α − Σᵢ Nᵢ eᵢα R_α}; setting (2) to zero gives R_α = max{0, K_α − Σᵢ Nᵢ eᵢα}, without the trailing R_α. (ii) The Jacobian (C1) has cᵢα Nᵢ in the resource–resource block where (2) gives eᵢα Nᵢ. (iii) App. C3 calls the Kaplan–Yorke dimension “the number of Lyapunov exponents that are positive” and then gives the correct interpolation formula (C4). (iv) In (E5) the error term’s sum runs over Tᵢ = noSS where it should run over Tᵢ = error; (E4) writes argmax = −argmin of the negative log-likelihood. None affects the phase boundary.
Open questions
- Does the boundary ρ* = √(S*/M*) survive structured (metabolic or networked) interactions, or externally supplied resources, which the authors report preliminary chaos for?
- Is there a replica-symmetry-broken description of the dynamic phase, as there is for random Lotka–Volterra models?