Represent each of m neural branches by a positive input field f_i and a positive output field g_i, then penalize violations of the paper's multi-output Borell-Brascamp-Lieb bound at weighted barycenters. The constraint couples branches through both local normalized ratios and global mass ratios, encouraging calibrated multi-view predictions without requiring all output functions to be identical.
Use the sharp exponential tail bound to set a local clipping threshold from a desired exceedance probability. Instead of globally clipping activations at a fixed value or percentile, clip each local window at its minimum plus B log(e/delta), where delta is the tolerated fraction of clipped entries.
Regularize hidden activations or attention logits by their local mean excess above the local minimum, rather than by symmetric variance or absolute magnitude. The penalty specifically suppresses upper-tail spikes while remaining invariant to adding a constant offset to every value in a local window.
Replace a learned causal mixing profile by a monotone profile followed by a prefix-average Hardy mixer. The monotonicity constraint makes the mixer provably non-degenerate in the BMO sense: localized variation in the profile cannot be reduced below a calibrated factor by prefix averaging. This is a cheap alternative to dense causal attention for tasks where importance or state profiles are expected to decay along sequence position.
Add a differentiable rearrangement-envelope penalty to a positive integral-kernel layer. The penalty uses the Laplace-transform inequality to prevent the sorted upper tail of the layer output from becoming substantially larger than the cumulative upper tail of its input, providing a distribution-sensitive alternative to ordinary activation clipping or an L2 penalty.
Build an attention or positive-mixture module whose output ratio at two control settings is provably monotone in an ordered index such as token distance, retrieval rank, or discretized uncertainty. Use normalized-positive-series identities to replace an unstable quotient derivative with a difference of expectations, and penalize violations of the resulting stochastic-order condition during training.
Generate temporal attention or convolution weights with the Graham–Knuth–Patashnik recurrence instead of learning every lag weight independently. For nonnegative recurrence parameters, the resulting lag sequence is strongly log-concave, so its normalized kernel is naturally unimodal and suppresses high-frequency sign-free oscillations without requiring a separate smoothness penalty. The six parameters can be learned per head, channel group, or layer, giving O(1) learned parameters for an…