Every idea extracted from recent arXiv mathematics papers — verified and unverified. Click an idea to open its full card; badges show the empirical verdict.
Replace a dense Gaussian random-feature layer by a sparse Poisson-chaos layer with the same learned kernel, and penalize the layer's normalized fourth add-one energy. The penalty suppresses rare single-atom jumps, so the sparse layer approaches the Gaussian same-kernel representation in Wasserstein distance while preserving computational advantages from sparse Poisson sampling.
Represent binary nuisance variables or augmentation bits as coordinates of a Hamming cube and penalize the model response that remains predictable from a random k-coordinate subset. Use the theorem's derivative-plus-global-norm certificate as the regularizer, retaining its p/log p dependence instead of using an arbitrary masking penalty. Random coordinate permutations provide a cheap stochastic approximation to the subset average.
Replace a dense directional feature map with a finite stable-zonotope layer: a small set of large random vectors creates a hyperplane arrangement over the input sphere, while a cheaper small-jump residual gives continuous features inside each cell. The layer exposes an explicit compute-versus-resolution knob through the jump cutoff and can be used either as an embedding before an MLP or as a lightweight router for mixture-of-experts blocks.
Maintain an SPD matrix preconditioner with the paper's deterministic Ornstein–Uhlenbeck covariance recursion rather than estimating an inverse through Newton–Schulz or an explicit matrix inverse. Apply this preconditioner to gradients from a small layer block, using damping and a conservative step size to preserve positive definiteness. The method is most plausible for low-rank, per-layer, or blockwise curvature matrices where dense matrix storage is affordable.
Paper: Theoretical Analysis of Thermodynamic Matrix Inversion: First-order Equivalence to Preconditioned Gradient Descent and Implications for Analog ComputingarXiv:2608.09743
Represent stochastic layer execution, branching, retries, and early exit as a finite continuous-time Markov chain, with the completed-prediction state absorbing. Learn transition rates jointly with neural-network weights, but use MFPT sensitivities to allocate rate changes according to their available control budget rather than allowing one routing edge to dominate halting-time control.
Build a sparse expert cascade whose router uses empirical progress, uncertainty, and compute cost to construct an efficient frontier of experts. Instead of always choosing the expert with the largest immediate gain per FLOP, route different confidence states through an ordered sequence of frontier experts, allowing cheap high-variance experts early and safer or more cost-effective experts near the final decision.
Add a response-sensitive regularizer to networks whose outputs should react predictably to a control input, using the stationary Markov sensitivity equation as a certificate. Instead of only penalizing large neural gradients, the method attributes amplification to the generator resolvent and can distinguish amplification caused by a nearly slow latent mode from amplification caused by uncontrolled parameter growth.
Use the paper's singular stopping-gain term to explicitly measure how much learned feature covariance crosses a max or routing boundary. Penalize excessive covariance in the normal direction to the switching surface, rather than pretending that the max operation has an ordinary Hessian.
Use spatially correlated training points whose low-frequency structure factor vanishes instead of iid points. For neural fields, PINNs, image-coordinate MLPs, or spatially indexed minibatches, this should suppress long-wavelength quadrature and gradient-estimation noise while preserving the represented target dynamics. The finite-order prediction is that a design with structure factor S(k)=O(|k|^{2q}) produces lower variance for smooth losses than iid sampling, especially as the domain or batch…
Replace ordinary additive or multiplicative activation noise with a nonnegative count-valued perturbation generated by the Hermite operator kernel. For a nonnegative feature x, sample an integer N whose distribution is exactly the operator's weight sequence and feed N/n to the next layer; the parameter alpha controls an additional even-jump component and therefore changes the noise geometry independently of the ordinary Poisson component.
Replace a fixed top-k expert count with a stochastic, token-specific fanout generated by a supercritical binary branching process stopped at a geometric time. The resulting number of active experts has finite mean but a power-law tail with log-periodic modulation, allowing most tokens to use little compute while reserving larger computation for difficult or ambiguous tokens.