Every idea extracted from recent arXiv mathematics papers — verified and unverified. Click an idea to open its full card; badges show the empirical verdict.
Parameterize a large linear layer as a sum of binary tensor products, W = Σ_l A_l ⊗ B_l, and regularize a factor-level upper bound on its top-k singular-value sums. The bound controls all Ky Fan norms of W while requiring SVDs only of the small factors, making it suitable for tensorized MLP or attention projections.
Replace a collection of overlapping sliding-window features with approximately orthogonal incremental features: the length-m feature contains information not predictable from shorter consecutive windows. Use the paper's transition-operator Toeplitz precision matrix to decorrelate the resulting sequence of window features before attention, suppressing duplicated local evidence and improving conditioning.
Instrument selected neural-network operators with cheap stochastic perturbations and estimate how much their outputs change under finite-precision perturbations. Use the resulting per-operator score to identify unstable kernels and selectively switch them to FP32, compensated accumulation, or a stable reformulation instead of running the entire model at high precision.
Build a low-rank trilinear attention module in which query, key, and value factors are constrained to the unit sphere and refined through a few proximal alternating sweeps. The proximal terms suppress factor oscillation and make each sweep improve a well-defined tensor interaction objective, offering a stable alternative to unconstrained tensor-power iterations.
Use the paper's effective-rank and eigengap-dependent covariance estimation rates to construct a confidence-aware low-rank bottleneck for transformer activations or key/value tensors. The bottleneck is enabled only when the top-p empirical eigenspace can be estimated more accurately than the desired compression error; otherwise the layer remains full-rank.
Construct sparse attention masks as regular or almost-regular hypergraphs: every query selects exactly w keys, while every key receives nearly equal incidence within each score or locality class. This preserves a fixed sparsity budget while removing systematic positional and head-level load imbalance that can cause hot keys, uneven memory traffic, and unstable gradients.
Add a learned scalar ordering to a directed graph attention layer and retain only forward edges, producing a DAG attention mask without requiring a supplied topological order. Train the ordering with a differentiable surrogate for weighted surplus, and regularize it toward the paper's explicit half-weight-minus-l2 certificate. This supplies a principled alternative to random masking or unconstrained bidirectional graph attention when causal or hierarchical information flow is desirable.
Compute separate doubled-angle orientation order parameters for left and right image regions, then expose their sum and difference as symmetric and antisymmetric global features. This gives a network a low-dimensional inductive bias for global vertical structure versus left-right imbalance, while retaining magnitude channels that indicate when either readout is undefined because orientations cancel.
Replace a dense block of a large neural-network weight matrix with an adaptively constructed cross approximation, but prevent premature termination using residual checks on O(m+n) additional diagonal-like entries. Accept a rank only after the probe residual has remained below tolerance for several consecutive iterations; otherwise continue adding pivots.
Construct attention groups as hyperedges of a linear r-uniform hypergraph: every pair of tokens is allowed to share at most one group, while each token participates in approximately the same number of groups. Apply local attention inside each group and aggregate the outputs across groups. The construction inherits the paper's sharp capacity bound and prevents both redundant pair interactions and high-degree token hubs.
Replace a dense channel-mixing matrix by a sparse complex generalised weighing matrix W with exactly w nonzero entries in every row and column, then use U=W divided by square root of w as a norm-preserving mixer. Restricting to k=2 gives a real matrix with entries in {+1,-1}; k=4 supports signed phase rotations. The exact isometry should preserve signal and gradient norms while reducing channel-mixing cost.
Replace an unconstrained relative-position attention-bias table by a learnable two-dimensional Fourier representation, or regularize the bias toward a small Fourier ratio. The ratio favors coefficient concentration without depending on the absolute scale of the bias, allowing the trained bias to be reconstructed from a small number of dominant frequencies.
Wrap a nominal gradient-based optimizer with a diagonal sign matrix that flips updates independently for parameter blocks, while a scheduler tests candidate sign configurations using short-horizon decrease of a Lyapunov-like training energy. The wrapper never changes the magnitude of the nominal update, and when the effective sign pattern is constant, it should recover the behavior of the correctly oriented nominal optimizer after a finite search period.
Assign separate sharpness or temperature parameters to two nonlinear subnetworks and anneal them according to a directional chart instead of driving both to their singular limits at the same rate. The optimizer explicitly tracks the ratio of the two scales and changes the schedule when the local Jacobian approaches a stability or bifurcation boundary. This tests whether the order and relative rate of sharpening, rather than only the final activation shape, controls optimization stability and…
Treat each neural-network block as a local strength system and measure how perturbations in its input channels affect multiple output observables, rather than using a single gradient norm. Use the estimated maximum directional gain to cap residual updates or assign a layerwise learning-rate multiplier, preventing weak high-gain layers from destabilizing training while allowing strong layers to move faster.
Replace expensive global spectral diagnostics of a cyclic or block-circulant neural layer by exact small Fourier-block calculations. Add a scale-normalized fourth-moment penalty, or directly cap the largest eigenvalue of each frequency block, to suppress frequency-specific amplification and reduce unstable training in long cyclic convolutions and structured attention.
Replace a dense token-mixing matrix in a sequence model with a fixed or learnable SBP derivative operator D=P^{-1}Q. The discrete integration-by-parts identity makes the interior mixing energy-neutral or boundary-dissipative, reducing exploding activations in deep residual stacks while preserving directional information along the sequence.
Replace activation-magnitude-based adaptive computation halting with a criterion based on the actual recurrent update and a local stability margin. The loop halts when the state change is small relative to state scale for several consecutive steps, avoiding pathological decisions when LayerNorm-driven dynamics cause the activation norm to collapse.
Decompose tensor-valued hidden states into invariant symmetric, alternating, and higher Young-symmetry channels before applying learned maps. This removes redundant tensor coordinates and prevents a neural layer from mixing incompatible representation types, yielding smaller equivariant modules with a cleaner inductive bias.
Add a targeted barrier or hinge loss to an existing attention or graph-mixing matrix that penalizes violations of signed circular-minor inequalities. Instead of enforcing only generic entrywise positivity, constrain higher-order noncrossing interactions encoded by determinants. This can suppress pathological oscillatory mixing while still allowing individual entries to be negative when the global structured sign pattern permits them.
Replace a fixed Fourier or spectral resolution in a neural operator or sequence model with a data-adaptive spectral cutoff. Keep only modes whose estimated signal energy exceeds the noise-amplification and discretization floor implied by the available number of trajectories and samples per trajectory. This should reduce overfitting to high-frequency sensor noise and preserve accuracy when the same model is deployed at different sampling resolutions.
Build a recurrent block as a fixed or learned ordering of local vertex foldings, mirroring the paper's identification of staircase solution maps with Coxeter elements of a folding group. Each folding changes one polygon coordinate by a rational cross-ratio completion while leaving all other coordinates unchanged. The resulting structured recurrence is reversible and can support constant-memory backpropagation by recomputing folds in reverse order.
Replace an unconstrained feature vector of size n+1 by the coefficients of a homogeneous degree-n binary polynomial and make the layer transform through the irreducible symmetric-power representation of GL_2(R). For n=4 this creates a five-channel equivariant feature block whose transformation law is exact rather than learned through augmentation.
Add a training-time regularizer that keeps the empirical joint covariance of hidden activations on multiple inputs close to the recursively predicted NNGP covariance. The regularizer targets the finite-width fluctuations quantified by the Wasserstein result, and is particularly appropriate for recurrent networks and attention blocks with shared weights, where hidden states at different positions or time steps are statistically coupled.