# One-Bang Gradient-Noise Preparation - ID: 836 - Canonical URL: https://synthcore.org/idea/836/one-bang-gradient-noise-preparation - API JSON: https://synthcore.org/api/idea/836.json - API Markdown: https://synthcore.org/api/idea/836.md - Verification status: beats_baseline - Source: [arXiv:2607.12955](https://arxiv.org/abs/2607.12955) - Category: dynamics - Solves: speed, stability, sample-efficiency - ML areas: optimizer, training-dynamics, scheduler - Math tags: control-theory, dynamical-systems, statistics, stochastic-processes, optimization - Ratings: usefulness 7/10; difficulty 5/10; novelty 7/10 ## Idea description Use a bounded stochasticity control during an initial preparation window to shape the gradient or parameter-update distribution before ordinary training. The control is restricted to its minimum or maximum value, with at most one switch, because the reduced moment dynamics are affine in the control; this gives a falsifiable alternative to smooth noise or learning-rate annealing. ## Mathematical statement For an update vector g in R^d, define its effective temperature T=E[||g||^2]/d and excess kurtosis a_2=(d/(d+2))E[||g||^4]/E[||g||^2]^2-1, paralleling the paper's kinetic state variables. Let u(t) in [u_min,u_max] be the controlled gradient-noise variance or learning-rate multiplier, and approximate the preparation dynamics by dT/dt=F(T,a_2)+B_T(T,a_2)u and da_2/dt=G(T,a_2)+B_a(T,a_2)u. For terminal objective J=phi(T(t_p),a_2(t_p)), Pontryagin's Hamiltonian is H=p_T[F+B_Tu]+p_a[G+B_au], where p_T and p_a are costates. Its control coefficient, or switching function, is S(t)=p_TB_T+p_aB_a. Minimizing J gives u=u_min when S>0 and u=u_max when S<0, with signs reversed for maximization. Thus an affine reduced model predicts a bang-bang schedule. Reachable states are obtained by integrating the two extreme controls; target kurtosis values outside the resulting interval are unattainable under the chosen optimizer and batch-noise amplitude. ## Key formulas - $$T=\frac{1}{d}\mathbb{E}[\|g\|^{2}],\qquad a_{2}=\frac{d}{d+2}\frac{\mathbb{E}[\|g\|^{4}]}{\mathbb{E}[\|g\|^{2}]^{2}}-1.$$ - $$\dot T=F(T,a_{2})+B_T(T,a_{2})u,\qquad \dot a_{2}=G(T,a_{2})+B_a(T,a_{2})u,\qquad u\in[u_{\min},u_{\max}].$$ - $$\mathcal{H}=p_T[F+B_Tu]+p_a[G+B_au],\qquad S(t)=\frac{\partial\mathcal{H}}{\partial u}=p_TB_T+p_aB_a.$$ - $$u^{\star}(t)=\begin{cases}u_{\min},&S(t)>0,\\u_{\max},&S(t)<0,\end{cases}\qquad \text{with at most one switch predicted during preparation.}$$ ## Implementation notes Integrate this at the optimizer or scheduler level for the first t_p updates of training. Choose a bounded control u to represent injected Gaussian gradient noise, dropout probability, or a multiplicative learning-rate/noise scale; the cleanest MVP is g_t=\widehat g_t+\sqrt{u_t}\epsilon_t, with epsilon_t sampled from a standard normal distribution and u_t in [u_min,u_max]. Maintain exponential moving averages of m_2=||g_t||^2 and m_4=||g_t||^4, then estimate T=m_2/d and a_2=(d/(d+2))m_4/m_2^2-1. First fit the local reduced dynamics by finite differences: run short constant-control probes at u_min and u_max, regress Delta(T,a_2) on (T,a_2,u), and estimate F,G,B_T,B_a. Specify a terminal target, such as minimizing a_2 at fixed T or maximizing the subsequent loss-decay rate over the next K updates. Numerically integrate the fitted two-dimensional model under both extreme controls, use the better endpoint, and implement a single-switch schedule: u_max for k