← Course

The 33 Families

From integer arithmetic to noncommutative geometry. A tour of every algebraic family that has demonstrated computational character — not just the seventeen the original v1 stack covered, but the sixteen more added in the eight-wave survey.

~5 minutes

What defines a family

A family is a set of operations that share the same algebraic structure — the same laws, the same identities, the same composition rules. Operations within a family can be combined and optimized together. Operations across families cannot.

The original Compute Stack v1 named seventeen families. The 2026 mathematics survey added sixteen more — bringing the total to thirty-three. The seventeen below are the foundations; we'll mention the sixteen new ones at the end.

The hardware-backed families

FOUND — Foundational algebra

Integer arithmetic. Bitwise logic. Floating point. Fused multiply-add. Comparison. Sorting. Quantization. Stochastic rounding. Interval arithmetic. Geometric algebra. The things your CPU already does.

LA — Linear algebra

Dense matrix multiply. Sparse matrix-vector products. Attention. The family that dominates modern AI. Every TPU, every GPU, every NPU optimizes this family first.

TR — Transforms and reductions

Reductions. Prefix sums. FFT. Convolution. Wavelet transforms. Pooling. Automatic differentiation. Multipole expansion. Radon transforms. The family of signal processing and structured reduction.

NL — Nonlinear

Activation functions. Softmax. Normalization. The operations that neural networks need but don't fit into linear algebra. All L₂ — all paying erasure cost.

MEM — Memory and state

Scatter and gather. Embedding lookup. Finite automaton steps. Relational joins. Constraint propagation. Unification. The operations that remember.

The hybrid families

FF — Finite fields and cryptography

GF(2ⁿ) arithmetic. Number theoretic transform (NTT). Polynomial ring operations. FHE bootstrapping. Modular exponentiation. Elliptic curve point ops. Lattice reduction. Bilinear pairings. The family that secures the internet — and the family most underserved by shipping silicon.

PROB — Probabilistic and stochastic

Hash functions. Random number generation. Monte Carlo accept/reject. MCMC / Langevin steps. The operations that turn determinism into probability.

CODE — Information coding

Syndrome decoding. Entropy coding (arithmetic and ANS). The operations that compress and error-correct.

The emerging families

SR — Semirings

Tropical semiring. Log-semiring. Viterbi semiring. Sum-product message passing. Fuzzy semiring. The family of dynamic programming, shortest paths, probabilistic inference, graph algorithms. Zero shipping silicon for any of them.

HD — Hyperdimensional

FHRR bind/unbind. Binary HD bundling and permutation. The family that represents meaning as 10,000-bit vectors and computes with them. Every operation is L₀ — free — and none of them have dedicated hardware.

TEN — Tensor and quantum

Tensor network contraction. Quantum gate operations. The mathematical family that generalizes matmul into higher dimensions and quantum superposition.

NEURO — Neuromorphic

Leaky integrate-and-fire (LIF) neurons. Spike-timing-dependent plasticity (STDP). The family that implements biological computation. Intel's Loihi and IBM's TrueNorth are the only serious silicon attempts.

ANALOG — Analog and physical

Analog crossbar MAC. Photonic MZI transforms. Sigma-delta modulation. B-spline / NURBS evaluation. The family that exploits physics directly instead of digitizing first.

STOCH — Stochastic and molecular

Gillespie steps. DNA strand displacement. The family where computation is chemistry.

The theoretical frontier

SYM — Symmetry and groups

Möbius/zeta transforms. Non-abelian group Fourier transforms. Symplectic integrators. The family that exploits symmetry for free computation.

OPT — Optimization

Proximal operators. Linear program pivots. The family of search and constraint satisfaction.

FRONT — Theoretical and formal

Reversible arithmetic. SAT/SMT oracles. Symbolic term rewriting. Discrete log oracles. The family at the edge of what's computable.

The sixteen new families

After the v1 stack was published, an eight-wave survey of every major branch of mathematics turned up sixteen more families — places where real computational operations exist but had never been formalized as primitives. None of them have dedicated silicon.

  • TOPO — topological operations: simplicial complexes, persistent homology, knot invariants, Steenrod squares, cyclic cohomology
  • CAT — category theory: functor application, adjunction resolution, subobject classifier evaluation
  • CALC — calculus and analysis: differentiation, integration, FEM, residue extraction, analytic continuation
  • STAT — statistical inference: MLE, Bayesian update, PCA, KDE, bootstrap, RANSAC
  • COMB — combinatorics: generating functions, Möbius inversion, matroid ops, RSK correspondence, Burnside enumeration
  • TYPE — type theory: beta reduction, type inference
  • STPROC — stochastic processes: Brownian increments, Kalman filters, particle filters
  • GEOM — computational geometry: convex hull, Voronoi, ICP, marching cubes, BVH traversal
  • GAME — game theory: Nash equilibrium, minimax, mechanism design
  • NT — number theory: primality, factoring, Hensel lifting, AKS, Pollard rho
  • DIFFGEO — differential geometry: parallel transport, exponential map, Dirac operator, Ricci flow
  • LANG — formal languages: parsing, NFA→DFA, BWT/FM-index, KMP, edit distance
  • ALGGEOM — algebraic geometry: Gröbner basis, sheaf cohomology, intersection multiplicity, Chow product
  • PHYSMATH — mathematical physics: partition functions, Feynman diagrams, lattice gauge sweeps, fermion matrix inversion, RG flow
  • DYNSYS — dynamical systems: ODE integration, Lyapunov exponents, bifurcation detection, attractor characterization
  • MEAS — measure theory: Lebesgue integration, Radon-Nikodym, Dixmier trace

The count

Thirty-three families. Two hundred fifty-eight primitives. Only about ten percent of them have dedicated silicon anywhere on Earth. Twenty-five entire families have zero hardware. The rest run on general-purpose chips that weren't designed for them, burning far more energy than physics requires.

Key takeaway

The industry has built extensive hardware for five or six families — FOUND, LA, TR, NL, MEM — and mostly ignored the other twenty-eight. Every primitive in those ignored families is either a research opportunity or a business opportunity. The stack shows you which.