The claim
Let components i = 1, …, k have abundances, or unnormalized participation intensities, n_i ≥ 0 and niche or feature vectors u_i ∈ ℝᵐ, the columns of U. Let b ∈ ℝᵐ be the environmental supply or target signal and c_i ≥ 0 a maintenance cost. Assume
A = UᵀU, r = Uᵀb − c, ṅ_i = n_i (r_i − (An)_i).
Define L(n) = ½‖b − Un‖² + cᵀn. Then:
- Gradient form. ∇L = UᵀUn − Uᵀb + c = An − r, so ṅ_i = −n_i ∂_i L. In logarithmic coordinates, d/dt log n_i = −∂_i L, which is continuous-time mirror descent with the entropy potential Σ n_i log n_i on the positive orthant.
- Descent. dL/dt = Σ ∂_i L · ṅ_i = −Σ n_i (∂_i L)² ≤ 0.
- Invasion fitness. A component’s per-capita growth is r_i − (An)_i = −∂_i L. At a resident equilibrium with n_i = 0, the rare component invades exactly when s_i = −∂_i L > 0: when increasing its weight from zero lowers the objective to first order.
- KKT. At an equilibrium every active component has ∂_i L = 0. If no absent component can invade, every absent one has ∂_i L ≥ 0. These are the KKT conditions for min L over n ≥ 0. Since A is positive semidefinite, L is convex, so such points are global minima. With A singular there may be several, and which one the flow reaches depends on the start.
So in this model three quantities coincide:
| ecology | optimization | learning |
|---|---|---|
| invasion fitness s_i | negative directional derivative −∂_i L at n_i = 0 | value of adding feature i |
It was checked numerically. For a random U (6 × 4), b and c, the identity ṅ = −n ⊙ ∇L holds to machine precision. Integrating the flow, L decreased at every step, and at the endpoint active components had zero gradient and the absent one a positive gradient.
The two-component case (CASE-118). With A = [[1, ρ], [ρ, 1]], 0 ≤ ρ < 1, and r₁ = r₂ = 1, the invasion growth of either component into the other’s equilibrium (n = 1) is 1 − ρ. That is also λ_min(A). The condition number of the Gram matrix is κ(A) = (1 + ρ)/(1 − ρ). In this model, niche overlap weakens the rare component’s invasion advantage and ill-conditions the feature system by the same parameter. The correspondence is exact here and does not extend as an identity to arbitrary competitive systems.
Why the record holds it
QUESTION-083 asks whether processes sharing one update rule can come to hold differentiated, mutually regulated authority. CLAIM-279 says roles are regimes, not kinds, and its note of 2026-10-18 brings in Chesson’s coexistence mechanisms. This claim gives the formal bridge those entries lacked. In the potential case, ecological coexistence of differentiated components is the same mathematics as sparse feature selection, and the recovery test of QUESTION-083 is a marginal-value test. It also marks exactly where the bridge ends (below).
Two boundaries of application
Abundance is not a normalized share of authority. The theorem is about nonnegative, unconstrained abundances. Influence in an organization is often relational, budgeted or competitive, and normalized so that Σ w_i = 1. That is a different dynamical model. If the shares are read off the abundances, w_i = n_i / Σ_j n_j, then with f_i(n) = −∂_i L(n):
ẇ_i = w_i ( f_i(n) − Σ_j w_j f_j(n) ).
This is a replicator equation, not the gradient flow on the orthant. If instead fitness is evaluated directly at w on the simplex, the equilibrium conditions carry a common Lagrange multiplier: ∂_i L(w) = λ for every active component, not ∂_i L = 0. So the equivalence between ecology and optimization does not automatically survive a change from abundance dynamics to conserved authority shares. Normalized endorsement is a separate model, and this claim does not cover it. That is not a claim that normalized endorsement cannot optimize. Normalization changes the feasible geometry and the equilibrium conditions, but it does not necessarily remove a potential-function reading. Whether one survives depends on how fitness and authority are defined on the simplex. For example, replicator dynamics whose fitness is evaluated at the shares themselves, f_i(w) = −∂_i L(w), are a gradient flow of −L on the simplex under the Shahshahani metric. The equation above, with fitness evaluated at the unnormalized n, is not of that form in general. (A point from the owner’s review of the filing, relayed 2026-10-18.)
The algebra holds for any U; the ecological reading needs a sign condition. The gradient-flow correspondence holds for arbitrary real feature vectors, because UᵀU is positive semidefinite. But signed feature vectors can have negative inner products. Then some off-diagonal entries of A are negative, and the corresponding Lotka–Volterra interaction is facilitative, not competitive. Reading A as a matrix of competitive niche overlaps also needs nonnegative off-diagonal interactions, as in models whose resource-use vectors are nonnegative.
What it does not say
- That ecological selection optimizes learning objectives in general. The identity needs symmetric interactions, A = UᵀU, and growth rates of the form Uᵀb − c. With asymmetric or nonreciprocal interactions (THEORY-634, Blumenthal, Rocks and Mehta 2024) there may be no scalar L at all.
- That operators with competing objectives are optimizers. An organization whose processes disagree about the desired outcome has, in general, no shared L whose gradient describes their interactions. That is where the equivalence breaks, and where it becomes interesting for agency. A local influence field might be split as F(n) = −∇V(n) + J(n), with J the non-potential part, but the split depends on a chosen metric and convention and is not unique without one. Whether organizations with the same local incentives and more non-reciprocity converge, oscillate, switch among configurations or become chaotic is open (THEORY-548 on delay-induced Lotka–Volterra instability, THEORY-634 on non-reciprocal catalytic rings).
- That the kernel fixes coexistence. The same Gram matrix supports coexistence or exclusion depending on r (CASE-118).
- That a population of coexisting specialists governs itself. Coexistence is not endogenous regulation (CLAIM-288).
Note of 2026-10-29: the cited duality and its neighbours, read
Seven works read on this date bear on this claim. None finds an error in the derivation.
- The duality, and its direction. Mehta, Cui, Wang and Marsland, LIT-1231 (NOTE-1395), confirm the summary’s citation. There the species are multipliers of constraints in resource space, and the Lotka–Volterra form with α_ij = Σ c_iα c_jα and κ_i = Σ K_α c_iα − m_i is the dual. That is this claim’s A = UᵀU and r = Uᵀb − c, with U = Cᵀ, b = K and c = m, so the claim’s objective is a dual objective.
- A third boundary: the reduction from resources. Marsland, Cui and Mehta, LIT-1230 (NOTE-1352): read as the fast-resource reduction of a consumer–resource model, A = UᵀU and r = Uᵀb − c hold only while every resource stays positive. Once a resource is driven extinct its row drops out, and the effective objective is MacArthur’s piecewise Q. The claim states the Lotka–Volterra form directly, so it is not wrong, but this belongs beside the two boundaries above. The same reading gives the objective a physical meaning: the perturbation the active components impose on the supply.
- The continuum case. Chizat and Bach, LIT-1232 (NOTE-1367): this claim’s invasion fitness, −∂_i L, is their −F′(µ)(u_i) restricted to a finite dictionary, and its KKT points are their Proposition 3.1’s minimizers on a fixed support. With positions free, the absent components form a continuum, and gradient flow transports mass without creating it, so a population reaches the global optimum only if a seed of every possible component is present from the start.
- A slowly evolving U. Power et al., LIT-1235 (NOTE-1368), keep the interaction matrix symmetric and rely on the potential that gives. Evolution there reshapes the objective the community descends, which the claim’s fixed-U setting does not cover.
- Symmetry is not uniqueness. Bunin, LIT-1234 (NOTE-1397): for symmetric interactions the random Lotka–Volterra model has MacArthur’s Lyapunov function, and its phase of multiple attractors still exists. So symmetry and an optimization principle do not by themselves give a unique outcome. That is consistent with this claim, whose A = UᵀU is positive semidefinite, so that L is convex.
- The asymmetric case. Blumenthal, Rocks and Mehta, LIT-1236 (NOTE-1366, THEORY-1162), give A = CEᵀ with C ≠ E and show what fails for a related consumer–resource model. It is the natural objection thread to the claim’s scope, not to the claim, which states its symmetric assumption. “What it does not say” cites this paper as THEORY-634, whose source is LIT-036, a different paper on non-reciprocal catalytic cycles; the reading of Blumenthal et al. is THEORY-1162.
- An open question, not a finding. Hofbauer and Schreiber, LIT-1233 (NOTE-1379), decide permanence by signs only when the invasion graph is acyclic. In this claim’s class L decreases along every trajectory, which should rule out heteroclinic cycles and so cyclic invasion graphs. The paper does not state this and the reading did not prove it.
The claim is unchanged.
Note of 2026-10-29: the invasion graph is acyclic, and when assembly has one end-point
The open point the earlier note of this date leaves, whether this claim’s class rules out cyclic invasion graphs, is closed by derivation. THEORY-1168 (Active), from Serván and Allesina, LIT-1824 (NOTE-1422): for any symmetric interaction matrix, the difference in rᵀx between two equilibria is a sum of invasion growth rates weighted by abundances, which an edge of Hofbauer and Schreiber’s invasion graph (LIT-1233) makes positive. So rᵀx rises along every edge and the graph has no cycles. Only symmetry and an equilibrium for each vertex are needed. In this claim’s class V(x*) = ‖b‖² − 2L(x*), so L falls along every edge, the class always meets LIT-1233’s acyclicity hypothesis, and the signs of invasion growth rates decide robust permanence there.
On order of arrival, the chapter gives one end-point, reached in any order of arrival, when the matrix is positive definite, which for A = UᵀU means U of full column rank. When A is singular, which the claim allows (“With A singular there may be several”), the chapter’s genericity assumption fails and it says nothing. Its indefinite case, acyclic but with priority effects, cannot arise from a Gram matrix. Two further limits on the transfer: the chapter needs positive entries in its matrix, which is this claim’s sign condition for the ecological reading, and its pool admits only species with r_i > 0.
The claim is unchanged.