quat-numpy 0.2.0

Quaternion algebra library with vector/matrix/tensor collections, linear algebra (SVD), interpolation (SLERP/squad), signal processing (QFFT/QConv), and optimized einsum kernels.

quat

Quaternion Algebra for NumPy

---

quat brings first-class quaternion support to NumPy. Scalar quaternions, vector/matrix/tensor collections, SVD, SLERP interpolation, QFFT signal processing, serialization — with a three-tier-einsum Hamilton kernel that picks the fastest strategy at every data size. All in pure Python, zero C extensions.

pip install quat-numpy

Why quat?

Most quaternion libraries stop at rotating 3D vectors. quat gives you the full toolbox:

What	How
Scalar quaternions	Full algebra — +, *, /, exp, log, pow, Hamilton product
Collections	QuatVector (1D), QuatMatrix (2D), QuatTensor (3D) — broadcast-aware
Linear algebra	SVD, rank, condition number, pseudo-inverse, determinant — on quaternion matrices
Interpolation	SLERP (shortest arc on S³), squad cubic spline, batch SLERP
Signal processing	1D/2D QFFT, quaternion convolution, FIR filter design
Random generation	Reproducible generators for all four types
Serialization	JSON + compact binary roundtrip for all types; SciPy Rotation interop
Performance	Three-tier Hamilton kernel (component-wise → einsum), ~20x faster SVD fast-path

Quickstart

import numpy as np
from quat import Quaternion, QuatVector, QuatMatrix
from quat import slerp, squad, random_unit_quat

# ---- Construction -------------------------------------------------
q = Quaternion(1, 2, 3, 4)                         # a + bi + cj + dk
r = Quaternion(5.0)                                 # pure real
u = Quaternion.from_axis_angle((0,0,1), np.pi/2)    # 90 deg around z
e = Quaternion.from_euler((0.1, 0.2, 0.3))          # intrinsic ZYX

# ---- Hamilton product ---------------------------------------------
i, j, k = Quaternion(0,1,0,0), Quaternion(0,0,1,0), Quaternion(0,0,0,1)
print(i * j)           # Quaternion(0, 0, 0, 1)   — i*j = k
print(i * j * k)       # Quaternion(-1, 0, 0, 0)  — i² = j² = k² = ijk = -1

# ---- Algebra ------------------------------------------------------
q.conjugate()          # Quaternion(1, -2, -3, -4)
q.norm()               # Euclidean norm
q.inverse() * q        # ≈ identity
q.exp()                # quaternion exponential
q.log()                # quaternion logarithm

# ---- 3D rotation --------------------------------------------------
u = Quaternion.from_axis_angle((0, 0, 1), np.pi / 2)
u.rotate_vector((1, 0, 0))           # ≈ (0, 1, 0) — 90° rotation
axis, angle = u.to_axis_angle()      # roundtrip (axis, angle)

# ---- Smooth interpolation -----------------------------------------
a = Quaternion(1, 0, 0, 0)
b = Quaternion(0, 1, 0, 0)
mid = slerp(a, b, 0.5)               # shortest arc on the 3-sphere

q0, q1, q2, q3 = [random_unit_quat() for _ in range(4)]
curve = squad(q0, q1, q2, q3, 0.75)  # cubic spline (Shoemake 1987)

# ---- Collections --------------------------------------------------
v = QuatVector([
    Quaternion(1,0,0,0), Quaternion(0,1,0,0), Quaternion(0,0,1,0)
])
v.real          # array([1., 0., 0.])
v.data.shape    # (3, 4)

A = QuatMatrix.eye(3)
B = QuatMatrix(np.random.randn(3, 4, 4))
C = A @ B       # quaternion matrix multiply (A * B also works)
C.shape         # (3, 4)

# ---- Linear algebra -----------------------------------------------
from quat.linalg import svd, pseudo_inverse, solve

U, s, Vh = svd(C)
A_pinv = pseudo_inverse(C)
x = solve(A_pinv, QuatVector(np.ones((3, 4))))

# ---- Signal processing --------------------------------------------
from quat.signal import qfft, qconv, lowpass

X = qfft(np.random.randn(256, 4))            # 1D quaternion FFT
k = lowpass(16, cutoff=0.2)                  # FIR lowpass filter
y = qconv(np.random.randn(128, 4), k._data)  # quaternion convolution

# ---- Serialization ------------------------------------------------
s = q.to_json()         # → '{"type":"Quaternion","data":[...]}'
b = q.to_bytes()        # → compact binary
q2 = Quaternion.from_json(s)
q3 = Quaternion.from_bytes(b)

# ---- Basis constants ----------------------------------------------
from quat import _I, _J, _K, _R, _ZERO
print(_I * _J * _K)     # -1 (Hamilton's fundamental identity)

API Reference

Quaternion — scalar quaternion

category	methods
constructors	zero(), one_q(), from_axis_angle(), from_euler(), from_complex_matrix(), from_real_matrix_left()
components	.r, .i, .j, .k, .real, .imag, .components, .data
arithmetic	+, -, *, /, -q, @
algebra	.conjugate(), .norm(), .normalize(), .inverse(), .exp(), .log(), .pow(t), .re_inner(q), .commutator(q)
rotation	.rotate_vector(v), .to_axis_angle(), .to_euler(seq='zyx')
validation	.isnan(), .isinf(), .isfinite(), .isclose(q)
serialization	.to_json(), .from_json(s), .to_bytes(), .from_bytes(b)
matrices	.to_complex_matrix(), .to_real_matrix_left(), .to_real_matrix_right()
conversion	float(q), int(q), complex(q), abs(q), np.asarray(q)

QuatVector / QuatMatrix / QuatTensor

	QuatVector	QuatMatrix	QuatTensor
ndim	1D	2D	3D
shape	(n,)	(m, n)	(p, q, r)
create	zeros, ones	zeros, eye	zeros
access	.data, .real/.i/.j/.k	.data, .real/.i/.j/.k, .row(i), .col(j)	.data, .real/.i/.j/.k
ops	.inner(v), .norm()	.norm(), .T, .H, .conjugate()	.inner(T), .unfold(mode), .mode_n_product(A)

Module overview

module	exports
quat.random	random_quat, random_unit_quat, random_quat_vector, random_quat_matrix, random_quat_tensor
quat.interpolate	slerp, slerp_vector, squad
quat.linalg	svd, rank, condition_number, pseudo_inverse, trace, det, norm, solve
quat.serialization	to_json, from_json, to_bytes, from_bytes, to_scipy_rotation, from_scipy_rotation
quat.signal	qfft, iqfft, qfft2, iqfft2, qconv, qconv2, lowpass, highpass, bandpass, bandstop
quat.optimized	hamilton_einsum, quat_matmul, conjugate_batch, norm_squared_batch, normalize_batch

Performance

The Hamilton product kernel uses a three-tier dispatch that selects the fastest strategy based on data size:

size	kernel	strategy
≤ 500 elements	_hamilton_component	direct float arithmetic
500 – 5000	_hamilton_einsum_noopt	einsum without contraction-path optimisation
> 5000	_hamilton_einsum	einsum with full optimisation

SVD fast-path: rank(), condition_number(), and norm(A, 2) compute only singular values via the complex (2×2) representation — ~20× faster than the full 4m×4n real SVD.

Zero-copy interop: q.to_numpy(copy=False) returns the internal buffer without copying. .to_complex_matrix() and .to_real_matrix_left() use np.empty to skip zero-initialisation.

Install

pip install quat-numpy

Requires Python ≥ 3.9, NumPy ≥ 1.21. Optional scipy for to_scipy_rotation / from_scipy_rotation.

For development:

git clone https://github.com/gealachlee/quat.git
cd quat
pip install -e .
pytest tests/