omnibias.graph
Spectral graph operators and differentiable relaxations for ordering, assignment, and graph-structured learning.
import omnibias.graphUse graph for Laplacian-based spectral operations or documented continuous relaxations of graph-related decisions. A graph adjacency matrix is not an arbitrary rectangular feature batch: check symmetry, weights, degree conventions, and the node axis for each operation.
Normalized Laplacians need a convention for zero-degree nodes. Spectral embeddings can be nonunique at repeated eigenvalues, so compare the relevant subspace or invariant quantity rather than assuming an identical eigenvector basis. Heat-kernel computations also depend on time and spectral scaling.
Gumbel, Sinkhorn, sorting, and top-k relaxations have temperature, iteration, and random-key or sampling contracts. A relaxed assignment is not automatically an exact permutation or optimum. Validate the decoded object and downstream loss separately. This alpha package uses AGPL-or-commercial licensing.
Module contract & mathematical conventions
Differentiable spectral graph operators and continuous relaxations.
PyTorch/JAX kernels provide Laplacians, spectral embeddings, graph heat kernels, Sinkhorn assignments, soft sorting, and top-k. Spectral algebra is exact for the supplied matrix up to numerical error; soft assignment is a relaxation. Discrete decisions and their optimality certificates belong to consumers.
Use matching float64 inputs and enable jax_enable_x64 for backend parity.
Framework reductions can differ in rounding; identical bits require matching
activation values and operation order, not merely shared coefficients.
Shared formulas do not imply universal floating-point identity. Verify dtype, shape, derivative-order support, and the stated numerical assumptions for the path you use.