# Multiplicative Log-Time Update Clock

- ID: 1398
- Canonical URL: https://synthcore.org/idea/1398/multiplicative-log-time-update-clock
- API JSON: https://synthcore.org/api/idea/1398.json
- API Markdown: https://synthcore.org/api/idea/1398.md
- Verification status: unverified
- Source: [arXiv:2607.25374](https://arxiv.org/abs/2607.25374)
- Category: dynamics
- Solves: stability, speedup, sample-efficiency
- ML areas: optimizer, training-dynamics, rnn, ssm, inference-speedup
- Math tags: stochastic-processes, statistics, dynamical-systems, probability
- Ratings: usefulness 6/10; difficulty 4/10; novelty 8/10

## Idea description

Use a multiplicative renewal clock to decide when a neural module is updated, rather than updating at every wall-clock tick or using a fixed iteration schedule. The resulting computation allocates many updates early and increasingly long intervals between later updates, while preserving a tunable stochastic distribution of update times; this is intended for anytime recurrent refinement, continual learning, or adaptive inference where late updates have diminishing marginal value.

## Mathematical statement

Let t_0>0 be the initial wall-clock time, let U_1,U_2,... be iid positive logarithmic inter-event variables with mean mu=E[U] and variance sigma^2=Var(U), and define T_n=t_0 exp(S_n), S_n=sum from i=1 to n of U_i. The number of updates by time t is N(t)=max{n:T_n<=t}, which is an ordinary renewal count in logarithmic time x=ln(t/t_0). Renewal asymptotics give E[N(t)] approximately x/mu and Var[N(t)] approximately (sigma^2/mu^3)x for large x. Thus the measurable signature is linear growth against ln(t/t_0), not against t, with Fano factor Var[N]/E[N] approximately sigma^2/mu^2. If a refinement state h is changed only at events, the neural dynamics are h(t)=h_{N(t)}, a random time change with logarithmically growing event count.

## Key formulas

- $$T_{n+1}=T_n\exp(U_{n+1}),\qquad T_n=t_0\exp\left(\sum_{i=1}^{n}U_i\right),\qquad U_i\overset{\mathrm{iid}}\sim p_U(u),\quad u>0.$$
- $$N(t)=\max\{n:T_n\le t\},\qquad \mathbb{E}[N(t)]\simeq\frac{\ln(t/t_0)}{\mu},\qquad \operatorname{Var}[N(t)]\simeq\frac{\sigma^2}{\mu^3}\ln(t/t_0),$$
- $$\frac{\operatorname{Var}[N(t)]}{\mathbb{E}[N(t)]}\simeq\frac{\sigma^2}{\mu^2},\qquad \frac{\sqrt{\operatorname{Var}[N(t)]}}{\mathbb{E}[N(t)]}\simeq\frac{\sigma}{\sqrt{\mu\ln(t/t_0)}}.$$
- $$h(t)=h_{N(t)},\qquad h_{n+1}=F_\theta(h_n,z_n),\qquad y(t)=G_\theta(h_{N(t)}).$$

## Implementation notes

Integrate the clock at the outer loop of an anytime recurrent-refinement model, state-space model, or continual-learning learner. Keep the neural transition h_{n+1}=F_theta(h_n,z_n) unchanged, but invoke it only when wall-clock time t reaches the next event time. At initialization set T=t_0, sample or estimate mu and sigma^2 from a chosen positive multiplier law p_U, and run: `T=t0; n=0; h=h0; while wall_time < budget: U=sample_pU(); T=T*exp(U); wait_until(T); h=F(theta,h,input); n=n+1; output=G(theta,h)`. For a deterministic ablation use U=mu; for a stochastic version use the paper's exact p_U if available, or fit a positive distribution to reported multiplier samples while preserving its first two moments. The paper supplies the renewal construction and asymptotic coefficients; the implementation estimates mu, sigma^2, event-time histograms, and task-specific output changes empirically. First run a cheap experiment with a GRU or four-layer recurrent refinement MLP on sequential MNIST or a small denoising task. Compare fixed-rate updates, geometric deterministic updates, and stochastic multiplicative updates at equal wall-clock budgets. Record N(t), validation loss, and hidden-state changes. The quantitative prediction is that E[N(t)] versus ln(t/t_0) is linear with slope 1/mu, while E[N(t)] versus t is sublinear. Var[N(t)] should have slope sigma^2/mu^3 against log-time, within about 20 percent after the renewal transient. Relative count scatter should decay as 1/sqrt(ln(t/t_0)). Falsify the idea if these slopes are absent or if it provides no anytime-compute advantage over a matched deterministic geometric schedule.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.