Linear algebra for the DeepCausality crates.
This crate owns the workspace's matrix representations and the algorithms over them. The split
against [deep_causality_tensor] is by arity, not density: a two-index object is a matrix
and lives here; ein_sum, broadcast, kronecker, the axis reductions and the tensor-train
stack take an N-index object and stay there.
Three representations, side by side
Choosing a representation is the central decision in linear algebra, and this crate holds all three so that the choice is local rather than architectural:
| representation | for |
|---|---|
| compressed sparse row | boundary and coboundary operators, discrete Laplacians |
| dense row-major | covariance, metric tensors, density matrices |
| bit-packed 𝔽₂ | mod-2 elimination, homology read as a code |
plus a dense vector, which the construction census found to be the larger need: 60 rank-1 constructions against 46 rank-2 across the seven consumer crates.
Banded on the tower, not on f64
Every operation is bounded on the weakest trait from deep_causality_algebra that makes it
correct, so the crate is generic over the integers as well as the floats. The determinant is a
polynomial in the entries and needs no division, so it is defined over any commutative ring —
bounding it on Field is what would have excluded ℤ. Gaussian elimination divides by its pivot
and so leaves ℤ on the first step, and is bounded accordingly.
The crate defines no scalar trait, marker or newtype of its own.
Position in the tier graph
This crate sits below deep_causality_tensor and must not depend on it under any feature.
CausalTensor's decompositions delegate into here, and a crate can only call into what it
depends on. Given that edge, the orphan rule leaves deep_causality_tensor as the only place the
access-trait impl for CausalTensor can be written.