sr-opinf 0.1.5

SR-OpInf: Symmetry-Reduced Operator Inference for Shift-Equivariant Systems

SROpInf — Symmetry-Reduced Operator Inference

SROpInf is a Python package for symmetry-reduced model reduction of shift-equivariant systems via operator inference. For a 1D periodic PDE whose solutions are dominated by a travelling / drifting structure (e.g. a travelling wave), it builds a reduced-order model (ROM) that first factors out the continuous translation symmetry and then learns the reduced dynamics in the co-moving frame — either intrusively (symmetry-reduced POD–Galerkin projection) or non-intrusively from snapshot data (symmetry-reduced operator inference, S-R OpInf).

The package accompanies the paper

Yu Shuai and Clarence W. Rowley, Symmetry-reduced model reduction of shift-equivariant systems via operator inference, arXiv:2507.18780 (2025). Submitted to Advances in Computational Mathematics (in revision).

and reproduces all of its Kuramoto–Sivashinsky (KS) numerical experiments.

---

Method at a glance

For a shift-equivariant system the solution q(x, t) is decomposed into a shift amount c(t) (the location of the travelling structure) and a co-moving profile obtained by aligning each snapshot to a fixed template q_template = cos(2πx/L) (template fitting). The reduced model then evolves

• the reduced co-moving state z(t) (POD coefficients of the template-fitted, scaled snapshots), and
• the scalar shift amount c(t) via a reconstruction equation dc/dt = numerator / denominator.

Two ways to obtain the reduced operators are provided:

ROM	Intrusive?	How the reduced operators are obtained
S-R POD–Galerkin	yes	Galerkin projection of the (scaled, symmetry-reduced) FOM right-hand side onto the POD subspace.
S-R OpInf	no	Ridge-regression of reduced polynomial operators to the data; optional re-projection for consistency and penalty (Tikhonov) regularization for stability.

Two spatial discretizations are supported, which lets the shift-induced interpolation error be assessed directly:

• Grid1DUniformSpectral — equispaced spectral grid; the shift and spatial derivatives are exact (Fourier), with 3/2-rule dealiasing for the quadratic nonlinearity.
• Grid1DCubicSpline — grid-based discretization; the shift, derivative, and inner-product operators are realized via periodic cubic-spline interpolation (hence an additional interpolation error, as encountered when non-intrusiveness is actually relevant).

---

Repository structure

SROpInf/
├── pyproject.toml              # package metadata + dependencies (installable, src-layout)
├── requirements.txt            # runtime deps (see also pyproject.toml)
├── requirements-dev.txt        # dev/test deps
├── RELEASING.md                # how to publish a new version to PyPI
├── LICENSE                     # MIT License
├── .github/workflows/
│   └── publish.yml             # CI: build + publish to PyPI on a v* tag (Trusted Publishing)
├── src/SROpInf/
│   ├── grids/grid1d.py         # Grid1D base + Grid1DUniformSpectral + Grid1DCubicSpline
│   │                           #   (inner product, sqrt-mass map R, shift_x, diff_x, fft/ifft)
│   ├── models/
│   │   ├── model.py            # FullOrderModel, SymmetryReducedScaledFullOrderModel (.project),
│   │   │                       #   SymmetryReducedScaledReducedOrderModel (.solve, .sample_and_compare)
│   │   └── ks.py               # KuramotoSivashinsky equation (polynomial operators)
│   ├── sr_tools.py             # template_fitting (shift amount c(t)); sropinf (the S-R OpInf
│   │                           #   regression, with re-projection / grid-search / cross-validation)
│   ├── mode_decomposition.py   # pod (energy-truncated POD of the symmetry-reduced snapshots)
│   ├── timestepper.py          # explicit (Euler/RK2/RK4) and semi-implicit (RK2CN/RK3CN) steppers
│   ├── dataloader.py           # FOMDataloader (loads trajectories / shift amounts from disk)
│   └── typing.py               # array type aliases
├── example/ks/
│   ├── configs.py              # all KS parameters and file paths (single source of truth)
│   ├── plot.py                 # figure helpers (x–t contours, spatial profiles, error bands, ...)
│   ├── ks_base_solution.ipynb  # experiment 1: single base travelling-wave solution
│   └── ks_perturbed_solutions.ipynb  # experiment 2: family of perturbed solutions (+ amplitude sweep)
└── test/                       # pytest suite (grids + model polynomial composition)

output/ (generated trajectories, data, and figures) is not tracked — it is fully regenerated by running the example notebooks.

---

Installation

Requires Python ≥ 3.10 (developed and tested with Python 3.13).

Which option do you need?

• Just using the library — if you only want to call the SROpInf functions from src/SROpInf in your own code, install from PyPI; that is all you need.
• Reproducing the paper experiments — the example/ notebooks are not shipped on PyPI (only the SROpInf library is). Clone the repository to get them, then install the library — see From source below.

From PyPI (recommended for library use)

python3 -m pip install sr-opinf

The package is then importable as SROpInf:

import SROpInf

The install name and the import name differ on purpose: you install the distribution sr-opinf but import the package SROpInf — the same pattern as pip install scikit-learn / import sklearn.

From source (for reproducing the experiments / development)

The example/ notebooks live only in the repository (they are not part of the PyPI package), so clone it first, then install the library:

git clone https://github.com/Yu-Shuai-PU/sr-opinf
cd sr-opinf
python3 -m pip install -e ".[examples]"   # editable install + Jupyter stack to run the notebooks
# extras: ".[examples]" -> jupyter + ipykernel + ipywidgets;  ".[dev]" -> pytest + ipywidgets
# or just: python3 -m pip install -e .     # runtime deps only (numpy, scipy, matplotlib, tqdm)

The -e (editable) install puts the SROpInf package on the path so the example notebooks can import SROpInf... directly. (You could instead pip install sr-opinf for the library — the key point is that running the experiments requires the cloned example/ files either way.) Exact dependency versions are declared in pyproject.toml.

---

Reproducing the paper experiments

The experiments live in example/ks/. Run the notebooks top-to-bottom:

1. Base travelling-wave solution — ks_base_solution.ipynb

Reconstruction (train = test) of a single KS travelling-wave solution. Builds the symmetry-reduced POD subspace and compares S-R POD–Galerkin against S-R OpInf (plain, re-projected, and penalty-regularized), on both the spectral and the cubic-spline grids — quantifying the effect of the shift-induced interpolation error.

First run is self-contained, just slower. The benchmark IC must sit on the KS attractor. If a cached output/ks_base_solution/data/traj_init_base.npy is not found, the notebook automatically runs a one-off t = 120 burn-in to generate (and cache) it before the T = 10 reconstruction — no manual configuration needed. Later runs reuse the cached IC and start immediately.

2. Perturbed solutions — ks_perturbed_solutions.ipynb

Prerequisite: run ks_base_solution.ipynb first (it generates the cached base IC this notebook reuses), and switch example/ks/configs.py to the perturbed-solutions case (type_traj_training and the matching base_path; the notebook's first cell asserts both).

Generalization across a family of solutions: trains on perturbed initial conditions and evaluates on a disjoint, held-out testing set. Includes a perturbation-amplitude sweep: per-trajectory testing errors are written to output/<case>/sweep_rRMSE/ and the final cell aggregates them into a box-plot of testing error vs. perturbation amplitude.

Configuration

All parameters and paths are centralized in example/ks/configs.py:

• type_traj_training — "base_solution" or "perturbed_solutions".
• physical / discretization — Lx = 2π, nx = 256, nu = 4/87, T = 10, dt = 1e-3.
• ROM — poly_comp = [1, 2] (linear + quadratic), pod_energy_threshold, the OpInf penalty_weight_* (Tikhonov penalties on the inferred operators), and shift_speed_denom_threshold (regularizes the shift-reconstruction denominator).

The output directory defaults to output/ks_multiple_solutions/ and can be overridden with the NISRH_OUTPUT_DIR environment variable. Each notebook's first cell asserts that configs points at the matching output folder and trajectory type, to prevent accidentally mixing cases.

---

Tests

python3 -m pytest            # from the SROpInf/ directory

The suite checks the grid operators (spectral vs. cubic-spline shift / derivative, inner products) and the polynomial-composition machinery of the model class.

---

Citation

If you use this code, please cite the accompanying paper:

@misc{shuai2025symmetryreduced,
  title         = {Symmetry-reduced model reduction of shift-equivariant systems via operator inference},
  author        = {Shuai, Yu and Rowley, Clarence W.},
  year          = {2025},
  eprint        = {2507.18780},
  archivePrefix = {arXiv},
  primaryClass  = {math.NA},
  note          = {Submitted to Advances in Computational Mathematics (in revision)},
  url           = {https://arxiv.org/abs/2507.18780}
}

License

Released under the MIT License (see LICENSE).