Galaga — Geometric Algebra for Python

Galaga 2 is a numeric geometric-algebra library built on a native Gram-matrix core. Its public API composes eager multivector values with optional immutable expression provenance, configurable blade conventions, and shared ASCII, Unicode, and LaTeX rendering.

• General symmetric Gram matrices, not only diagonal signatures.
• Long, unambiguous operation names as the primary mathematical API.
• Optional concise aliases such as gp, op, rev, and sw.
• Explicit inner products and contractions rather than one overloaded ip.
• Immutable names and expression provenance over eager numeric values.
• Presets that configure the metric, blade convention, notation, and model metadata together while still permitting component-level overrides.
• Thread- and async-safe scoped presentation changes.
• Read-only coefficient access and checked scalar conversion.

Install

For the Galaga 2 prerelease train:

python -m pip install --pre "galaga>=2.0.0a1,<3"

After stable 2.0.0 is published:

python -m pip install "galaga>=2,<3"

Galaga requires Python 3.11 or newer and NumPy. The optional galaga-anywidget visualization package also supports Python 3.11, while galaga-marimo requires Python 3.14 because it uses t-strings.

Quick start

from galaga import Algebra, DisplayPolicy, norm, outer_product

alg = Algebra((1, 1, 1), display=DisplayPolicy("full"))
e1, e2, e3 = alg.basis_vectors(expr=True)

u = (2 * e1 + e2).named("u")
v = (0.5 * e1 + 1.8 * e2).named("v")
area = outer_product(u, v)

print(area)
print(area.latex())
print(norm(area))

The operators are conveniences over named operations:

u * v  # geometric_product(u, v)
u ^ v  # outer_product(u, v)
u | v  # doran_lasenby_inner(u, v)
~u     # reverse(u)

Constructing an algebra

Diagonal algebras retain the familiar call shapes:

from galaga import Algebra

cl3 = Algebra(3)                       # Cl(3, 0)
sta = Algebra(1, 3)                    # Cl(1, 3)
pga = Algebra(3, 0, 1)                 # Cl(3, 0, 1)
ordered = Algebra((1, -1, -1, -1))    # explicit basis order

A full symmetric Gram matrix defines a non-orthogonal or native-null basis:

import numpy as np
from galaga import Algebra

null_plane = Algebra(
    gram=np.array(
        [
            [0.0, -1.0],
            [-1.0, 0.0],
        ]
    )
)

Complete presets configure numeric and presentation choices together:

from galaga import Algebra, p_cga, p_lengyel_cga, p_pga, p_rga, p_sta

sta = Algebra(config=p_sta("mostly-minus"))
pga = Algebra(config=p_pga(spatial_dim=3))
cga = Algebra(config=p_cga(spatial_dim=3, frame="null"))
lengyel_cga = Algebra(config=p_lengyel_cga())
rga = Algebra(config=p_rga(spatial_dim=3))

config= owns the whole algebra definition. The lower-level presentation=, blades=, notation=, local_names=, display_order=, and display= parameters permit deliberate overrides when constructing an algebra directly.

Native-null conformal model

The conformal preset stores eo and einf as actual null basis vectors in a non-diagonal Gram matrix. Attach the model-specific semantics explicitly:

from galaga import Algebra, outer_product, p_cga
from galaga.cga import ConformalModel

algebra = Algebra(config=p_cga(spatial_dim=3))
cga = ConformalModel(algebra, expr=True)
expanded_cga = cga.with_expression_form("expanded")

a = cga.up((0, 0, 0))
b = cga.up((1, 0, 0))
line = outer_product(a, b, cga.infinity)

assert cga.carrier(cga.round_point((1, 2, 3))).homogeneous_grade() == 2

round_point = cga.round_point((3, 4, 0), radius_squared=4)
assert float(cga.center_norm(round_point)) == 5
assert float(cga.radius_norm(round_point)) == 2

# The same value can explain itself with compact CGA vocabulary or its formula.
compact = cga.carrier(round_point)
expanded = expanded_cga.carrier(round_point)
assert compact == expanded

The native-null CGA guide covers round and flat objects, operator/expanded expression forms, att/car/ccr/cen/con/par, dual conventions, projection, Eric Lengyel's ●/○/■/□ components and weighted norms, and transformation recipes.

Point-based rigid model

p_rga() selects Eric Lengyel's point-based interpretation of Cl(3, 0, 1). Attach RigidModel when coordinates, projective measurement, projection, support, or validity constraints are needed:

from galaga import Algebra, p_rga
from galaga.rga import RigidModel

algebra = Algebra(config=p_rga())
rga = RigidModel(algebra, expr=True)

p = rga.point((3, 4, 0)).named("P")
q = rga.point((1, 0, 0)).named("Q")
line = p ^ q

assert float(rga.bulk_norm(p)) == 5
assert rga.is_valid_line(line)

The RGA guide covers the algebraic convention layer, paired norms, homogeneous distance and angle, projections, support, line/motor/flector constraints, transwedge correction, and the dual relationship with plane-based PGA.

Values, names, and expressions

Multivectors are eager and immutable. Naming and expression tracking return a new wrapper without changing the coefficients or the original value:

from galaga import Algebra, exp

alg = Algebra(3)
e1, e2, _ = alg.basis_vectors(expr=True)

theta = alg.scalar(0.6).named(r"\theta")
B = (e1 ^ e2).named("B")
R = exp(-theta * B / 2).named("R")

R.name       # immutable semantic Name
R.expr       # immutable expression provenance
R.numeric    # presentation-independent galaga.core.Multivector
R.data       # read-only coefficient array

Useful state transformations include:

R.named("Q")       # replace the semantic name
R.without_name()   # retain expression provenance
R.with_expr()      # attach literal provenance if absent
R.without_expr()   # retain the eager value and name

Factories accept expr=True when the construction itself should participate in a later expression tree:

e1, e2, e3 = alg.basis_vectors(expr=True)
scalar = alg.scalar(2, expr=True)
vector = alg.vector([1, 2, 3], expr=True)

Blades

blade() accepts configured labels, bitmasks, signed blade references, or an existing signed unit blade. blades() is the ordered plural form:

rga = Algebra(config=p_rga())
e1, e2, e3, e4 = rga.basis_vectors(expr=True)
e23, e31, e41, e42 = rga.blades(
    e2 ^ e3,
    e3 ^ e1,
    e4 ^ e1,
    e4 ^ e2,
    expr=True,
)

alg.locals() returns a read-only mapping for environments where bulk name injection is appropriate. In reactive notebooks, explicit blade() or blades() calls preserve dependency tracking more clearly.

Generated conventions support compact, juxtaposed, and wedge blade products:

Model-aware convention builders retain semantic roles while changing that spelling. Their dimension argument is the model's spatial dimension, not the total algebra dimension:

from galaga import Algebra, null_cga_blade_convention, p_cga

cga = Algebra(
    config=p_cga(spatial_dim=3),
    blades=null_cga_blade_convention(3, style="juxtapose"),
)

This is a five-dimensional algebra whose pseudoscalar renders as $e_1 e_2 e_3 e_o e_\infty$. The convention adds the origin and infinity vectors to the three Euclidean vectors itself.

Product and contraction family

The long names are canonical:

There is deliberately no top-level ip or inner_product: these operations disagree for mixed grades and scalar inputs. Choose the intended definition explicitly, or define a local notation:

from galaga import doran_lasenby_inner as ip

Variadic geometric and outer products lower left-to-right through their binary catalog operation:

from galaga import geometric_product, outer_product

geometric_product(a, b, c)
outer_product(e1, e2, e3)

Commutators and anticommutators are unscaled:

commutator(a, b)            # ab - ba
anticommutator(a, b)        # ab + ba
half_commutator(a, b)       # (ab - ba) / 2
half_anticommutator(a, b)   # (ab + ba) / 2
lie_bracket(a, b)           # unscaled commutator
jordan_product(a, b)        # unscaled anticommutator

Grades and scalar conversion

from galaga import grade, grades, scalar_part

scalar_component = grade(value, 0)      # scalar multivector
selected = grades(value, [0, 2, 4])
coefficient = float(grade(value, 0))
same_coefficient = scalar_part(value)   # optional helper

float(value) succeeds only when the entire multivector is scalar. It never silently discards non-scalar grades. The explicit .data property exposes the read-only NumPy coefficient array:

coefficients = value.data

Multivectors deliberately do not implement NumPy's array or ufunc protocols; np.asarray(value) is not a coefficient conversion.

Rendering and presentation

DisplayPolicy chooses content independently from the target format:

from galaga import Algebra, DisplayPolicy

alg = Algebra(3, display=DisplayPolicy(content="full"))
e1, e2, _ = alg.basis_vectors(expr=True)
x = (2 * e1 + e2).named("x")

x.display("full/latex")
x.display("expr/unicode")
x.display("value/ascii")
x.latex()

Persistent presentation changes return cheap algebra views sharing the same numeric algebra. Scoped overrides use ContextVar, so they are isolated by thread and asynchronous task:

teaching_alg = alg.with_presentation(teaching_presentation)

with alg.use_presentation(teaching_presentation):
    print(x.latex())

Public API and numeric core

Import application APIs from galaga. Internally, those public objects are owned by the composition facade:

import galaga
import galaga.facade

assert galaga.Algebra is galaga.facade.Algebra
assert galaga.Multivector is galaga.facade.Multivector

Use galaga.core when presentation, names, and expression provenance are not needed:

from galaga.core import Algebra, geometric_product

numeric = Algebra(gram=[[1.0, 0.2], [0.2, 1.0]])
e1, e2 = numeric.basis_vectors()
result = geometric_product(e1, e2)

Prerelease legacy oracle

During the Galaga 2 prerelease migration, Galaga 1 remains deliberately available as an isolated test oracle:

from galaga import legacy

old_alg = legacy.Algebra(3)
old_e1, old_e2, _ = old_alg.basis_vectors()
old_value = legacy.gp(old_e1, old_e2)

Legacy and Galaga 2 values are separate domains and must not be mixed. The galaga.legacy namespace, including galaga.legacy.render and galaga.legacy.simplify, is migration infrastructure and is scheduled for removal with the old table engine before stable 2.0.0.

More documentation

• Documentation index
• Galaga 1 to 2 migration guide
• Numeric core
• Galaga 2 implementation overview
• Native-null CGA
• Rigid Geometric Algebra
• Release process

Executable examples live in the repository gallery.