Three-Equilibrium Reaction RNN Memory Cell
Implementation & benchmark of arXiv:2608.01116 — The Ubiquity of Three Steady States: Minimal Multistable Zero-One Reaction Networks
Source paper: The Ubiquity of Three Steady States: Minimal Multistable Zero-One Reaction Networks arXiv:2608.01116 ⓘ · analyzed Aug 31, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Replace a generic recurrent update by a positive-state continuous-time cell whose interactions are restricted to a quadratic zero-one reaction-network motif with three state variables and six reactions. Select a motif known to possess three positive equilibria, then use the two stable equilibria as binary memory states and the intervening unstable equilibrium as the separatrix. This creates an explicitly multistable RNN module with a bounded attractor count and a measurable stability certificate.
Formulas
Mathematical statement
Let x(t)=(x_1,x_2,x_3)^T in R_{>0}^3 be the cell state, let N in Z^{3x6} be the stoichiometric matrix, Y in {0,1}^{3x6} the reactant matrix, and k_j>0 the six reaction rates. For reaction j, define the mass-action monomial h_j(x)=product_i x_i^{Y_{ij}}, with sum_i Y_{ij}<=2, and let the neural dynamics be f(x)=N h(x)+u(t), where h=(k_1 h_1,...,k_6 h_6)^T and u is a bounded input injection. At a positive equilibrium z_l, f(z_l)=0 when u=0. The Jacobian is J(z_l)=N diag(h(z_l)) Y^T diag(z_l)^{-1}. For the paper's full-dimensional quadratic zero-one class, consecutive positive steady states satisfy sign det J(z_{l+1})=-sign det J(z_l), and the three-state motif has at most three positive equilibria; the bistable case therefore has stable-unstable-stable equilibria. Stability must be checked with the full cubic Hurwitz conditions, not determinant alone: for p(lambda)=lambda^3+a_1 lambda^2+a_2 lambda+a_3, asymptotic stability requires a_1>0, a_2>0, a_3>0, and a_1 a_2>a_3.
Implementation notes
1. Exact integration point: implement this as a continuous-time recurrent layer inside an RNN or latent world model. Choose one of the paper's 373 three-species, six-reaction motifs, or enumerate candidate integer matrices N and Y and retain only motifs with two stable and one unstable positive equilibria. Parameterize k_j=softplus(theta_j)+epsilon and x_i=softplus(q_i)+epsilon so positivity is preserved. Use an input encoder to produce u(t), with a small leak or bounded input clamp to prevent input-driven escape. Decode the state with a linear layer or small MLP.
2. Pseudocode: initialize x=softplus(q). At every solver step compute h[j]=k[j]*product(x[i]**Y[i,j]), compute dx=N@h+u, and integrate with RK4 or an adaptive ODE solver. During training, periodically solve f(x)=0 from many positive initializations, compute J=N@diag(h)@Y.T@diag(1/x), calculate its eigenvalues and Hurwitz margins, and add a penalty when the two intended memory equilibria violate the Hurwitz inequalities. Train the readout and rates jointly, but freeze the motif topology or perform a discrete motif search before gradient training.
3. Computed from the paper: the allowed quadratic zero-one topology, the six-reaction/full-rank restriction, the alternating determinant-sign rule, and the maximum of three positive equilibria. Estimated empirically: equilibrium locations, basin volumes, eigenvalues, and the parameter value at which a stable/unstable pair disappears. Do not use determinant sign as the sole stability test; verify all eigenvalues or the cubic Hurwitz conditions.
4. First cheap experiment: compare this cell with a three-unit tanh RNN on a two-bit associative-memory task and a delayed XOR sequence task. Use identical parameter counts and train on short sequences, then test recall after long free-running horizons under small perturbations. Sweep one rate multiplier k_1=c k_1^0 and use multistart root finding to measure the number of positive equilibria. The predicted signature is a stable-unstable-stable ordering with opposite signs of det(J) at consecutive roots, at most three positive roots, and a sharp bistability transition at a saddle-node: the smallest Jacobian eigenvalue should approach zero from below at the boundary, while basin occupancy changes sharply. A successful implementation should reproduce the equilibrium pattern across at least 90% of random positive initializations and locate the transition within 20% of a high-resolution parameter sweep.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.