Nonequilibrium Sensitivity Certificate
Implementation & benchmark of arXiv:2607.12096 — Topological building blocks of nonequilibrium response
Source paper: Topological building blocks of nonequilibrium response arXiv:2607.12096 ⓘ · analyzed Aug 30, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Add a response-sensitive regularizer to networks whose outputs should react predictably to a control input, using the stationary Markov sensitivity equation as a certificate. Instead of only penalizing large neural gradients, the method attributes amplification to the generator resolvent and can distinguish amplification caused by a nearly slow latent mode from amplification caused by uncontrolled parameter growth.
Formulas
Mathematical statement
Let x be a controlled input, W(x) a strongly connected continuous-time Markov generator, pi(x) its normalized stationary distribution, and g an output observable. The response derivative satisfies W pi' = -W' pi with 1^T pi' = 0. The group inverse W# acts on the zero-sum subspace and gives pi' = -W# W' pi. Therefore the response derivative R'(x) = g^T pi'(x) obeys the bound |R'(x)| <= ||g||_2 ||W#(x)||_2 ||W'(x) pi(x)||_2. Sensitivity increases when a nonzero generator eigenvalue approaches zero, because ||W#|| is approximately inverse to the smallest nonzero spectral gap. A practical certificate penalizes violations of a prescribed response bound B(x): L_cert = E_x[softplus(|R'(x)| - B(x)) squared].
Implementation notes
Attach the certificate to a neural module that predicts a scalar response from a controllable input, such as a classifier probability, dose-response model, or neural ODE with discrete latent modes. Parameterize a small generator W(x) from the hidden state, compute its stationary distribution pi by solving the augmented linear system formed by stacking W and a row of ones against a right-hand side containing zeros and one, and obtain pi' by solving the same augmented system with right-hand side formed by -W' pi and zero. Compute R' = g^T pi' and add the certificate penalty to the task loss. Pseudocode is: h = encoder(x); rates = softplus(Ah + b) + epsilon; construct W; solve for pi; differentiate rates to obtain W'; solve for pi'; Rprime = g^T pi'; loss = task_loss + lambda times relu(abs(Rprime) - B) squared. Compute W' with automatic differentiation and estimate the smallest nonzero eigenvalue of W for diagnostics. The stationary and sensitivity equations are mechanistic; B and lambda are empirical hyperparameters. First test a three-state latent generator on synthetic nonmonotonic curves against an unconstrained MLP while sweeping learning rate. The prediction is |R'| proportional to 1 divided by the spectral gap gamma when W' pi is fixed, and the certificate should keep response derivatives below B within 10 percent.
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.