RC-Mode Response Regularizer
Implementation & benchmark of arXiv:2609.02458 — Impedance in Periodically Driven Stochastic Systems
Source paper: Impedance in Periodically Driven Stochastic Systems arXiv:2609.02458 ⓘ · analyzed Sep 3, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Use the paper's parallel-branch response structure as a measurable regularizer on training dynamics. Penalize large high-frequency gain and excessively slow relaxation modes, encouraging parameter updates whose loss response is smooth, damped, and composed of controlled time scales rather than a sharp unstable mode.
Formulas
Mathematical statement
For stable relaxation modes with time constants \(\tau_k>0\), susceptibility is \(H(i\omega)=\sum_k a_k/(1+i\omega\tau_k)\). Each branch attenuates high frequencies with magnitude \(|a_k|/\sqrt{1+\omega^2\tau_k^2}\). Define an empirical response penalty \(\mathcal R_{\mathrm{RC}}=\sum_j w_j|\widehat H(i\omega_j)|^2+\alpha\sum_k[\log(\tau_k/\tau_0)]_+^2\), where \(\omega_j\) are probe frequencies, \(w_j\) are nonnegative weights, \(\tau_0\) is a target maximum time constant, \(\alpha\) controls slow-mode suppression, and \([q]_+=\max(q,0)\). The total objective is \(L_{\mathrm{total}}=L+\beta\mathcal R_{\mathrm{RC}}\), where \(L\) is task loss and \(\beta\) is regularization strength.
Implementation notes
Attach the regularizer to a network during training, evaluating it once every M steps to control overhead. Inject small sinusoidal perturbations into the learning rate or update magnitude using \(\theta_{t+1}=\theta_t-\eta_t[1+\epsilon\sin(\omega t)]g_t\). Use a normalized Gaussian direction if a parameter projection is needed, and measure the scalar response as training loss, validation loss, gradient norm, or projected displacement. Estimate \(\widehat H(i\omega_j)\) with lock-in Fourier correlations over at least four probe periods. For the first MVP, do not differentiate through the estimator: compute the response penalty after each probe window and use it to rescale subsequent updates. Fit positive \(\tau_k\) and residues when mode interpretation is useful; otherwise use measured squared amplitudes directly. The paper supplies the parallel relaxation-mode decomposition, while response amplitudes, probe frequencies, mode count, and target \(\tau_0\) are empirical. First experiment: train a small CNN on CIFAR-10, comparing cross-entropy alone with the response penalty at matched compute and several \(\beta\) values. The prediction is that high-frequency gain decreases monotonically with \(\beta\), oscillatory loss disappears near the fixed-learning-rate boundary, and robustness to a twofold learning-rate perturbation improves. Also test whether the dominant fitted pole remains stable and whether loss autocorrelation decays faster than in the unregularized baseline.
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.