linoss-dynamics

linoss-dynamics is a small NumPy runtime package for LinOSS-style oscillator stepping with explicit non-negative damping support.

Status: public alpha.

Vision

linoss-dynamics is the runtime physics layer for oscillatory state-space dynamics. The current scope (v0.2) covers the deterministic oscillator step, irregular-time stepping, Bayesian filtering, and stability/frequency introspection. The longer roadmap — including unified oscillator + attractor primitives and spatio-temporal extensions — is documented in docs/roadmap.md.

How this differs from the upstream LinOSS reference

The original tk-rusch/linoss is a training-time JAX/Equinox neural-network library that exposes LinOSS as trainable layers (LinOSSLayer, LinOSSBlock, LinOSS) inside a deep-learning model. It is the right choice when you are building or training a sequence model that uses LinOSS as a learnable encoder.

linoss-dynamics is a runtime physics package. It keeps the oscillator stepping idea and IM/IMEX discretization family, but presents them as small NumPy functions for deterministic runtime use rather than trainable neural network layers:

Who needs linoss-dynamics specifically

• Replay-safe agentic systems — needed inside deterministic kernels where every step must be byte-reproducible (e.g. consumed by elume at runtime).
• Simulation and control — robotics, signal processing, vibration analysis, and any context that uses oscillatory dynamics as a runtime model rather than a learnable layer.
• Reference implementation for porting — a clean NumPy baseline to validate custom JAX/PyTorch ports of LinOSS against.
• Embedded or minimal-dependency contexts — anywhere a JAX install is too heavy or unavailable.
• Teaching and pedagogy — a self-contained, well-tested, single-file solver that students can read end-to-end in an afternoon.
• Energy-based analysis — applications that need to monitor energy drift and detect convergence; these diagnostics are not exposed by the upstream reference.

If you are training a neural network, use the upstream package. If you are running a system that needs LinOSS as a deterministic physics step, use this one.

What It Provides

• Classic implicit and implicit-explicit LinOSS step helpers.
• Explicit non-negative damping G for D-LinOSS-style runtime damping.
• Energy, energy-delta, and convergence-window helpers.
• A dependency-light package core with NumPy as the only runtime dependency.

Install

From PyPI (recommended):

pip install linoss-dynamics

From GitHub (latest main):

python -m pip install "linoss-dynamics @ git+https://github.com/bionicbutterfly13/linoss-dynamics.git@main"

For local development:

git clone https://github.com/bionicbutterfly13/linoss-dynamics.git
cd linoss-dynamics
python -m pip install -e ".[test]"

Quickstart

Classic LinOSS-style step:

import numpy as np
from linoss_dynamics import linoss_step

y = np.array([0.2])
z = np.array([1.0])
A = np.array([1.0])

y_next, z_next, metrics = linoss_step(y, z, A, dt=0.1, mode="implicit")

print(y_next, z_next, metrics["energy_after"])

Damped step with explicit G:

import numpy as np
from linoss_dynamics import damped_linoss_step

y = np.array([0.2])
z = np.array([1.0])
A = np.array([1.0])
G = np.array([0.5])

y_next, z_next, metrics = damped_linoss_step(y, z, A, G, dt=0.1)

assert metrics["damping_mode"] == "explicit_g"

Sequence-level scan over a time series:

import numpy as np
from linoss_dynamics import linoss_scan

U = np.random.default_rng(0).standard_normal((100, 1))  # (T, n)
A = np.array([1.0])

Y, Z, metrics = linoss_scan(U, A, dt=0.05)
print(Y.shape)  # (101, 1) — includes initial state at Y[0]

Bayesian filtering over an observed time series (requires [probabilistic]):

# pip install linoss-dynamics[probabilistic]
import numpy as np
from linoss_dynamics import kalman_filter, rts_smoother, oscillator_mats

Ad, Qd, Bd = oscillator_mats(beta=0.1, omega=2.0, sigma_proc=0.3, dt=0.05)
H = np.array([[1.0, 0.0]])
R = np.array([[0.1]])

y_obs = np.sin(np.linspace(0, 10, 200)) + 0.1 * np.random.default_rng(42).standard_normal(200)
filt = kalman_filter(y_obs, Ad, Qd, H, R)
smooth = rts_smoother(Ad, filt)

print(filt["m_f"].shape)   # (200, 2)
print(smooth["m_s"].shape) # (200, 2)

Parameter fit on observed data (requires [probabilistic]):

# pip install linoss-dynamics[probabilistic]
import numpy as np
from linoss_dynamics import fit_oscillator_mle

y = np.sin(np.linspace(0, 20, 400)) + 0.15 * np.random.default_rng(0).standard_normal(400)
result = fit_oscillator_mle(y, dt=0.05)

print(f"fitted omega = {result['omega']:.3f} rad/s")
print(f"fitted period = {result['period']:.3f} s")

Runnable tutorials for all use cases are in examples/.

For a single bounded prompt/context packet covering the package purpose, symbol surface, LinOSS relationship, and adjacent technology categories, see docs/heredoc-context.md.

Numerical Contract

A controls oscillator stiffness/frequency.

G controls damping/forgetting.

The package deliberately keeps those controls separate. Damping is not modeled as hidden A scaling, and negative G values are rejected.

Supported damping shapes:

• scalar G
• vector G with the same length as y and z
• diagonal matrix G

Supported step modes:

• implicit / IM
• implicit_explicit / IMEX

Public API

Solver

Stability

Continuous (irregular-time)

Probabilistic (requires pip install linoss-dynamics[probabilistic])

Public errors:

Scope

The package does not implement JAX training loops, Discretax integrations, active-inference runtimes, metacognitive policy, event buses, graph databases, or web APIs.

Host applications should keep adapters outside this package and depend on linoss-dynamics through the public API above.

Development

Run the full local check set:

python -m ruff check src tests
python -m pytest tests -v --tb=short
python -m compileall -q src tests

See CONTRIBUTING.md, PROVENANCE.md, and CLAIMS.md before changing package behavior or public wording.

Citation

If this package is useful in research or technical writing, cite the package metadata in CITATION.cff and cite the upstream LinOSS and D-LinOSS papers listed below.

Release and publishing steps are tracked in docs/release-checklist.md.

Attribution

This package is not the original LinOSS or D-LinOSS research implementation.

• LinOSS / Oscillatory State-Space Models: T. Konstantin Rusch and Daniela Rus, Oscillatory State-Space Models, arXiv:2410.03943.
• D-LinOSS / learned damping: Jared Boyer, T. Konstantin Rusch, and Daniela Rus, Learning to Dissipate Energy in Oscillatory State-Space Models, arXiv:2505.12171.
• Official LinOSS ecosystem: https://github.com/tk-rusch/linoss.
• Discretax / former Linax ecosystem: https://github.com/camail-official/discretax.

See PROVENANCE.md and CLAIMS.md before making public claims.