# Pointwise complexity-gated inference

- ID: 3016
- Canonical URL: https://synthcore.org/idea/3016/pointwise-complexity-gated-inference
- API JSON: https://synthcore.org/api/idea/3016.json
- API Markdown: https://synthcore.org/api/idea/3016.md
- Verification status: unverified
- Source: [arXiv:2609.01576](https://arxiv.org/abs/2609.01576)
- Category: architecture
- Solves: inference-speedup, accuracy, stability
- ML areas: transformer, diffusion, inference-speedup
- Math tags: probability, stochastic-processes, metric-geometry, measure-theory
- Ratings: usefulness 7/10; difficulty 5/10; novelty 6/10

## Idea description

Use a local chaining complexity computed from an empirical input metric to predict stochastic output error for each individual input. Easy, locally concentrated inputs can use fewer dropout, ensemble, or diffusion samples, while high-complexity inputs receive additional computation; unlike a global confidence threshold, the allocation varies with the input.

## Mathematical statement

For an anchored Banach-valued stochastic process Z_t indexed by a separable space T, t_0 is an anchor, d is a pseudometric controlling sub-Weibull increments, v(t)=d(t,t_0), and μ is a reference probability measure on T. The pointwise Fernique-Talagrand functional of order α>0 is Φ_{μ,d}^{(α)}(t)=∫_0^{4v(t)}[log(1/μ(B_d(t,r)))]^{1/α}dr, where B_d(t,r)={s∈T:d(s,t)≤r}. The paper's simultaneous bound states that with probability at least 1−δ, for every t simultaneously, ||Z_t|| is at most a constant times Φ_{μ,d}^{(α)}(t)+v(t)[log(e/δ)]^{1/α}. The bound remains valid for a data-dependent selected index. In the adaptation, Z_x is stochastic network-output error, d(x,x') is an empirical robust output metric, μ is the validation distribution, α is a fitted tail exponent, and δ is the target failure probability.

## Key formulas

- $$\Phi_{\mu,d}^{(\alpha)}(t):=\int_{0}^{4v(t)}\left(\log\frac{1}{\mu(B_d(t,r))}\right)^{1/\alpha}dr,\qquad v(t):=d(t,t_0),\qquad B_d(t,r):=\{s:d(s,t)\le r\}.$$
- $$\mathbb{P}\left(\forall t\in T:\ \|Z_t\|\lesssim \Phi_{\mu,d}^{(\alpha)}(t)+v(t)\left(\log\frac{e}{\delta}\right)^{1/\alpha}\right)\ge 1-\delta.$$
- $$\widehat d(x,x')=\operatorname{Quantile}_{q}\left(\left\|f_{\xi_q}(x)-f_{\xi_q}(x')\right\|\right),\qquad \widehat\mu(B)=\frac{1}{n}\sum_{i=1}^{n}{\bf1}\{x_i\in B\},$$
- $$M_\delta(x)=C\left[\widehat\Phi_{\widehat\mu,\widehat d}^{(\alpha)}(x)+\widehat d(x,x_0)\left(\log\frac{e}{\delta}\right)^{1/\alpha}\right],\qquad K(x)=\min\{K_{\max},\max\{K_{\min},\lceil cM_\delta(x)^2/\varepsilon^2\rceil\}\}.$$

## Implementation notes

Integrate this at inference for a stochastic predictor: a dropout transformer, ensemble, stochastic-depth network, or diffusion model using repeated denoising trajectories. Let x be the input, f_ξ(x) the output under random seed ξ, and x_0 a validation-set medoid. Collect n calibration inputs and Q stochastic outputs per input. Estimate the robust empirical metric \widehat d(x,x') as a high quantile or median of ||f_{ξ_q}(x)−f_{ξ_q}(x')||, using logits or compact embeddings rather than full generated outputs. Approximate μ(B_d(x,r)) with k-nearest-neighbor counts in the calibration set. Evaluate the displayed Φ integral on a logarithmic radius grid from a small positive floor to 4\widehat d(x,x_0). Fit α and the multiplicative constant C on a held-out calibration split so M_δ covers observed high-sample output error at the desired rate. At inference, run K_min stochastic samples, estimate M_δ(x), then allocate K(x) additional samples; stop when the running mean's empirical confidence radius is below ε. The paper supplies the envelope structure, while metric estimation, C, α, and the mapping from M_δ to Monte Carlo error are empirical. First test a 10-member dropout ResNet-18 on CIFAR-10 and a small dropout language model. Compare fixed-K inference and global variance gating at equal error and equal compute. Success is lower average sample count at matched accuracy, empirical failure probability near δ, and positive correlation between M_δ(x) and actual stochastic prediction error.

## Disclaimer

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