Blog

What's New in Deep Causality 0.13.7

Marvin Hansen

This release touches 16 crates across the workspace. Four ship breaking changes: ultragraph (0.8 → 0.9), deep_causality_multivector (0.4 → 0.5), deep_causality_topology (0.5 → 0.6), and deep_causality_physics (0.5 → 0.6). The rest reamain fully API-compatible.

Five themes stand out.

1. Precision is now a parameter

The biggest cross-crate change. deep_causality_physics, deep_causality_sparse,deep_causality_multivector, deep_causality_tensor, and deep_causality_topology were generalized over R: RealField. Hard-coded f64 is gone from kernels, unit types, and state carriers. That includes astro, fluids, MHD, photonics, EM fields, relativity spacetime, dynamics, thermodynamics, nuclear, materials, quantum gates, and MaxwellSolver. HilbertState and HopfState are now generic over R with no f64 default.

In practice, you can run the same kernel with f32 for throughput, f64 for production, or extended-precision types such as Float106 when stability matters, all by Changing just one type alias. Take a look at the mathematics examples for how precision as a parameter is used in practice.

This is the breaking change behind the major bumps in physics, multivector, and topology crates. Migration is mechanical: add an R: RealField parameter, or pin a concrete float type at the call site.

2. Fluid dynamics, end to end

deep_causality_physics gained a complete fluid-dynamics module, added in 15 staged groups:

  • Governing equations: Navier–Stokes residual, conservation forms.
  • Constitutive kernels: viscous stress. ViscousStress and ReynoldsStress were promoted from a shared CauchyStress to dedicated newtypes.
  • Dimensionless numbers: Reynolds, Mach, Prandtl, and friends.
  • Turbulence: TKE, dissipation, eddy quantities.
  • Coherent-structure detectors: Q-criterion, λ₂, vorticity-based detectors.
  • Compressible-flow thermodynamics and boundary-layer kernels.
  • Regime evaluators: incompressible Newtonian NS, Stokes, Euler, and compressible Newtonian NS. Each is a distinct regime with its own validity checks.
  • Reference-solution verification tests for the NS regimes (Poiseuille, Couette, etc.).

Raw-scalar inputs in three fluid kernels are now validated rather than trusted, and the test tree was restructured to mirror src/kernels/.

3. Topology: Hodge decomposition and cubical Regge calculus

deep_causality_topology added real differential geometry.

Hodge decomposition lands as a first-class operation:

  • New HodgeDecomposition carrier type and matching error variant.
  • A hodge_decompose algorithm using matrix-free conjugate gradient.
  • A HasHodgeStar<R> capability trait, with simplicial and cubical implementations (UnitEdge, Uniform, PerAxis, and PerEdge tiers).
  • Property-based tests covering orthogonality and idempotence.
  • Lazy ⋆ population via OnceLock, so the star operator is built once and cached.

Cubical Regge calculus ships as a working core:

  • Cell volumes, hinges, dihedrals, deficit angles, and the discrete Einstein–Hilbert action on cubical lattices.
  • A Lorentzian signature marker and Wick-rotated action. Signature truth is routed through deep_causality_metric.
  • Regge action gradient plus a Metropolis–Hastings sampler, with a single-edge gradient on the hot path.
  • A Manifold generic widening over ChainComplex, so the same manifold types work simplicially or cubically.

Other topology additions:

  • TopologicalInvariants extractor (B1a). Betti numbers and friends in one pass.
  • PointCloud::triangulate_delaunay with discriminating errors for degenerate input, capped at ambient dimension.
  • LatticeComplex plus cubical aliases and Neighborhood strategies.
  • A new TopologyErrorEnum::HodgeDecompositionFailed variant.

The breaking changes here include the removal of the old Lattice, LatticeWitness, and DualLattice structs, plus the CWComplex trait. All four are superseded by the new LatticeComplex-based design.

4. Ultragraph 0.9: biconnectivity

ultragraph gained a biconnectivity decomposition:

  • articulation_points, bridges, and biconnected_components are now methods on StructuralGraphAlgorithms.
  • Benchmarks for the decomposition path.

5. A witness-typed isomorphism trait family

deep_causality_num got a three-tier isomorphism trait stack:

  • Tier 1: marker subtraits that declare “these two carriers are iso.”
  • Tier 2: witness-typed Iso traits with StandardIso as the canonical witness.
  • Tier 3: the integration layer, including From and witness round-trip checks in both directions.

Six concrete iso instances ride on top. One of them is a structural iso between CausalMultiField and its tuple carrier in deep_causality_multivector. Along the way, six pre-existing bugs blocking no_std were fixed in deep_causality_num.

Smaller things worth knowing

  • deep_causality_effects was retired; the generalize-effects-over-realfield spec was archived alongside it.
  • CI: codecov, Miri, and OSV GitHub Actions were updated. Several test modules now skip cleanly under Miri (CDL SURD, transcendental floats, rusty-fork modules, stat and float doctests).
  • deep_causality_core pinned Functor/Monad consistency between propagating-effect witnesses.
  • deep_causality_physics flattened its unit-type folder and consolidated examples into a dedicated examples folder.

Upgrading

If you only depend on the API-compatible crates (deep_causality, deep_causality_core, deep_causality_num, deep_causality_uncertain, deep_causality_tensor, deep_causality_algorithms, deep_causality_discovery, deep_causality_ethos, deep_causality_macros, deep_causality_sparse, deep_causality_rand, deep_causality_haft), this is a drop-in upgrade.

For the four breaking crates, the migration story is:

  • physics, multivector, topology: thread an R: RealField parameter through your code, or pin it to f64 at the boundary. Drop any <f64> turbofish; it is no longer accepted where the kernel is now generic.
  • topology: replace Lattice, DualLattice, and CWComplex usage with the new LatticeComplex API.
  • ultragraph: if you implement StructuralGraphAlgorithms yourself, add the three new methods. If you only call it, no action needed.

Full per-crate detail is in each crate’s CHANGELOG.md.