Chernoff-Tied Neural Evolution
Implementation & benchmark of arXiv:2609.02727 — Neural operators approximate strongly continuous convex monotone semigroups
Source paper: Neural operators approximate strongly continuous convex monotone semigroups arXiv:2609.02727 ⓘ · analyzed Sep 3, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Replace a conventional deep neural operator with repeated applications of one learned one-step operator whose parameters are shared across time. Train the block at a small step size and require its short-horizon compositions to match observed finite-time evolution, making depth correspond to physical or algorithmic time rather than an arbitrary number of layers.
Formulas
Mathematical statement
The paper constructs a semigroup from one-step operators using $S(t)f=\lim_{n\to\infty}I_n^{k_n^t}f$, where $S(t)$ is the target evolution semigroup, $f$ is an input function, $I_n$ is a one-step operator, $h_n>0$ is its step size with $h_n\to0$, and $k_n^t=\max\{k\in\mathbb{N}_0:kh_n\le t\}$ is the number of steps up to time $t$. The generator is $Af=I'(0)f=\lim_{n\to\infty}(I_nf-f)/h_n$, where $A$ is the infinitesimal generator. The transferable principle is that a stable, monotone, convex one-step approximation can be composed many times to approximate the full semigroup. Implement a neural approximation $\widehat I_h$ satisfying $\widehat I_h f\approx f+h\widehat A(f)$ for small $h$, and use tied parameters at every step. If the block has Lipschitz factor $L_h\le1+Ch$, then an initial one-step error $\varepsilon_h$ grows at most as $(1+Ch)^k\varepsilon_h\le e^{Ct}\varepsilon_h$ over $k\approx t/h$ steps.
Implementation notes
Integrate this at the entire neural-operator block, using an MLP, graph neural network, or discretized function encoder that predicts a function-valued derivative. Use one parameter set theta for every evolution step. Inputs are a discretized function f, a fixed step h, and an integer horizon K. Pseudocode: y = f; for k in range(K): z = A_theta(y, conditioning); y = y + h*z; return y. Train on pairs (f, target_at_t) with K=floor(t/h), and include a one-step loss ||A_theta(f)-(target_at_h-f)/h||^2 whenever short-time data are available. Also include a composition loss by applying the same block for K steps and comparing with the target at t. Estimate the empirical Lipschitz factor by drawing perturbations u and computing L_hat=max ||I_hat(y+u)-I_hat(y)||/||u||; apply spectral normalization or reduce h when L_hat > 1+C*h. The paper supplies the semigroup construction and generator scaling; Lipschitz constants, spectral norms, and generator targets are estimated in code. First experiment: learn Burgers or 2D advection-diffusion evolution with a small graph or Fourier-style operator, comparing ordinary untied depth-K layers against tied Chernoff blocks at equal parameter count and FLOPs. Measure rollout error versus horizon, stability under 5-10x longer rollout, and error at equal parameter count. Success is slower long-horizon error growth and equal-or-better accuracy with K times fewer parameters.
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.