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 fixed-norm Fourier feature layer by a Fourier transform followed by spatially varying modular normalization. Use a baseline exponent approaching the endpoint regime at large coordinates and permit only bounded, smooth deviations so the transform remains controlled while the network can emphasize localized details.
Replace an isotropic Fourier-feature map with a fractional low-pass map whose order is selected from the estimated intrinsic Frostman dimension of the training samples. The layer represents a coefficient vector f in the ambient domain, applies the multiplier |k|^{-s}, and evaluates the smoothed function on the observed fractal-like data support. The theorem provides a geometry-dependent bound preventing high-frequency coefficient energy from producing arbitrarily large responses on concentrated…
Treat batches of samples, modalities, or MoE experts as components of a differentiable mixture and add the paper's topology-sensitive RPA free energy to the training objective. Learn a low-dimensional topology descriptor for each component, map it to an effective structure factor, and use the resulting free energy either to promote specialization or to penalize unwanted phase separation in representations.
Periodically project a rectangular neural-network weight matrix onto an approximately orthonormal-column matrix using LU-preconditioned CholeskyQR rather than ordinary QR or a polar iteration. Pivoted LU handles badly scaled and nearly dependent columns, while Householder orthogonalization of the LU factor produces a triangular preconditioner that makes the subsequent Cholesky step safer in fp16 or bfloat16.
Add a local curvature penalty to graph learning or GNN training that penalizes sampled node signals with negative discrete Bakry–Émery curvature. The regularizer targets graph bottlenecks and irregular diffusion geometry, and can be applied either to a learned adjacency matrix or to the task-relevant hidden representations propagated by a fixed graph.
Build a linear state-space or recurrent layer in a learned pseudo-unitary coordinate frame $\Theta(t)$, and penalize the covariant coefficient $P_{m,\Theta}$ instead of penalizing $\Theta'(t)$ or transition-matrix norms directly. The regularizer is sensitive to meaningful variation of the represented Hamiltonian but is invariant to redundant gauge representations, potentially reducing unstable latent modes without forcing every parameter matrix to be small.
Augment a representation-learning objective with penalties enforcing the paper's four-point metric inequalities, and use an exponential snowflake kernel instead of unconstrained dot-product similarity. The experiment tests whether geometrically valid similarities improve retrieval or attention stability at equal model size and compute.
Represent the sequence of hidden states through a residual or state-space network as a polygonal curve and penalize turns according to their signed moment arm relative to the curve's input and output states. This targets bends that most strongly reduce endpoint separation, rather than applying an unweighted total-curvature penalty. The expected benefit is better long-range signal transport and less folding of hidden trajectories at comparable parameter count.
Use a two-gradient predictor-corrector average as the gradient supplied to Adam, retaining trajectory smoothing while avoiding the three or four gradient evaluations required by full RK3. Vary the mixing coefficient to test whether the reported regularization comes from gradient averaging itself rather than from high-order integration.
Regularize a scalar feature field on a 2D grid by interpreting each feature value as the uniformizing variable of a hyperbolic ring and penalizing violations of local orthogonal-ring angle closure. Unlike a raw Laplacian penalty, this constrains the representation through positive hyperbolic radii and geometrically meaningful edge compatibility.
Maintain a small population of neural-network parameter replicas and interleave ordinary gradient steps with Boltzmann/Kac-style binary collisions. Each collision preserves the pair's mean parameter vector and relative-distance norm while randomly rotating the relative direction, with collision frequency proportional to a regularized negative power of replica distance.
Construct a recurrent or state-space block with two learned transition matrices A and B representing two commuting update directions. Besides penalizing noncommutation and deviation from isometry, penalize the negative spectrum of the paper's core operator H(A,B), encouraging a structured overlap of one-step and two-step ranges. Compare this against an orthogonal-RNN baseline and against commutation-only regularization on long-horizon sequence tasks.
For a learned phase-space layer, estimate its symplectic Fourier bandwidth R and divide its output gain by the theorem's support-dependent factor R raised to an exponent determined by the Schatten index p. This creates a resolution-aware normalization: layers with larger phase-space bandwidth are automatically damped when p is not equal to 2, while the Hilbert-Schmidt case p = 2 remains unscaled.
Generate structured augmentations of categorical sequences using the paper's adjacent crystal rewrites, then enforce prediction consistency across the resulting orbit. Unlike arbitrary random swaps, the rewrite preserves paired subsequences and modifies only the unmatched portion, making it appropriate for exchangeable discrete codes or explicitly permutation-equivariant inputs.
Construct a shared latent transformation as a neural monad-like operator Γ=Ω∘Σ, and expose its iterates Γ^{q+1}Y as a refinement trajectory rather than stacking unrelated layers. Aggregate the resulting representations with a learned or fixed realization weighting, while training an algebra-action map θ:ΓY→Y to make one-step refinement compatible with the original representation. This creates a shallow-parameter, arbitrarily deep computation path with explicit compositional…
Replace homogeneous feature propagation with a discretized wave equation containing a positive, spatially varying learnable potential. The potential changes Hamiltonian trajectories so that feature energy reaches the layer's readout or sensor region instead of remaining in dynamically hidden modes. Train the potential jointly with the task objective and an empirical observability penalty.
Add a loss that prevents an intermediate feature map from being simultaneously concentrated inside a porous spatial region and a porous frequency region. The regularizer is based on the fractal uncertainty inequality: if frequency support is restricted to a porous set Y, then the fraction of feature energy inside a porous spatial set X is at most C h^beta; violations of this bound are penalized.
Use the derivative of fractional feature energy with respect to its order as a regularizer for intermediate representations. This penalizes unstable scale behavior rather than simply suppressing all high frequencies, so it can preserve useful detail while discouraging uncontrolled changes across spatial scales.
Use the paper's finite-habitat approximation as a warning and design principle: averaging token- or state-dependent routing environments can reduce the persistence of specialized subnetworks. Partition inputs into environments, estimate environment-specific interaction kernels, and retain the heterogeneity that produces positive invasion margins instead of replacing it with one global average.
Represent tokens, features, or attention states by normalized rank-one matrices and train the network to preserve their Schatten-p distance profiles over complex phase rotations. Because the paper proves that equality of all distances \(\|\lambda e-v\|_p\) identifies \({\rm Tr}(e^*v)\), this regularizer preserves matrix overlap geometry under a learned transformation.
Treat the learned latent transition F_theta as a homeomorphism-like operator and monitor the range of its temporal-difference operator D_theta u = u composed with F_theta minus u. If the smallest nontrivial singular values of the sampled operator collapse toward zero as trajectory length or basis size grows, the latent dynamics are entering an ill-conditioned coboundary regime. Use this signal to reduce the recurrent step size, impose contraction, or replace the transition by a periodicized…
Represent MoE experts as leaves of a balanced ternary tree and regularize the hierarchical boundary of each expert's assignment mask. At fixed routing mass x, the ternary martingale isoperimetric theorem supplies the explicit minimum one-variation T_3(x), so the router can be penalized according to an occupancy-dependent profile rather than a uniform parent-child disagreement cost. This should favor coherent, stable routing regions while preventing small expert supports from obtaining…
Build a variational quantum neural network whose trainable 2-qubit Hamiltonian is projected into the Lee-Yang coupling cone and augmented by a uniform field term -h sum_i Z_i. The theorem certifies a nondegenerate ground state and a gap at least h/4, enabling imaginary-time state-preparation layers with predictable exponential suppression of excited-state error.
Use a barycentric rational activation or filter whose interpolation nodes are periodically zoomed into the range of preactivations or eigenvalues actually encountered by the network. Protect the layer from catastrophic poles by monitoring the associated generalized eigenproblem and penalizing poles close to the active input interval. This targets rational networks whose expressivity comes from localized poles but whose training is destabilized by denominator zeros.