DDALAB

ddalab is the unified Python package for DDALAB.

It ships:

• the command-line interface
• the Qt desktop application
• the shared local runtime and dataset readers
• the bundled dda-rs backend binary used for local DDA analysis

Entry Points

Installing this package provides:

• ddalab
• ddalab-cli
• ddalab-gui

Use ddalab open to open the desktop application. Other ddalab commands use the command-line interface. ddalab gui remains an alias for ddalab open, and ddalab-gui launches the application directly.

Install

python -m pip install ddalab
ddalab open

For a source checkout:

cd packages/ddalab
python -m pip install .

Published wheels and desktop installers bundle the local dda-rs backend, so DDA works offline without a separate network service.

For editable or source installs, local DDA analysis also requires a working Rust toolchain with cargo so the bundled dda-rs binary can be built for your platform.

CLI Examples

ddalab health
ddalab dataset info --file data/MG100_Seizure1.edf
ddalab dda info --json
ddalab dda validate data/MG100_Seizure1.edf --json
ddalab dda run data/MG100_Seizure1.edf --channels 0 1 2 --variants ST SY --end 30
ddalab dda run data/MG100_Seizure1.edf --channels 0 1 2 --variants ST --device cuda
ddalab dda batch --bids-dir data/ds003029 --variants ST --continue-on-error

CPU execution is the default. The desktop DDA setup and CLI accept CUDA on systems with an NVIDIA driver and cuBLAS runtime; cuda:N selects a specific GPU. Analyses that use Rust-only CCD extensions continue those extension computations on the CPU.

Python DDA Estimators

The bundled Python backend includes scalar estimators for derivative-form and weak-form DDA. Derivative-form DDA builds delayed polynomial features Phi(t) and regresses a numerical derivative of x(t) on those features. Weak-form DDA uses the same delayed polynomial dictionary, but fits integrated increments instead:

x(t + Delta) - x(t) ~= integral_t^{t+Delta} Phi(s) ds @ beta

This avoids pointwise derivative estimation and can be preferable when observation noise would be amplified by finite differences. Coefficients remain on the same differential-equation scale as derivative-form DDA because both the response and design matrix are integrated over the same window.

from ddalab_app.backend.dda import DerivativeDDA, WeakFormDDA

vanilla = DerivativeDDA(degree=3, delays=[0], derivative="finite_difference")
vanilla.fit(x, dt=0.01)

weak = WeakFormDDA(
    degree=3,
    delays=[0],
    integration_window=9,
    stride=1,
    quadrature="trapezoid",
    regression="ridge",
    ridge_alpha=1e-6,
    standardize_features=True,
)
weak.fit(x, dt=0.01)
print(weak.term_names_)
print(weak.coefficients_)

integration_window and stride are specified in samples. Use integration_window_seconds or stride_seconds with dt/sample_rate when time-based settings are more convenient. The default weak-form window is conservative: half of the smallest positive delay, with a lower bound of three samples, or five samples for nondelayed scalar systems. Too-small windows behave more like derivative DDA; too-large windows can blur fast dynamics.

The available quadrature rules are trapezoid, left_rectangle, right_rectangle, midpoint, and simpson when SciPy is available. trapezoid is the default because it is the natural continuous-time weak-form choice for sampled trajectories. left_rectangle can match an Euler-stepped simulator more closely and is useful as a discretization diagnostic, but it is not the recommended default for continuous-time estimation.

Weak-form DDA does not universally dominate derivative DDA. It trades derivative noise sensitivity for integration-window bias. In practice, choose integration_window on a held-out validation interval by minimizing weak-form interval prediction error, then report the selected-window result. Oracle-best windows, selected using known synthetic coefficients, are useful for experiments but should be treated only as upper bounds.

For noisy or ill-conditioned windows, the recommended weak-form setting is regression="ridge" with standardize_features=True. Feature standardization is used only internally for numerical conditioning; exposed coefficients are back-transformed to the original feature scale, so they remain comparable to ordinary derivative-form DDA coefficients.

Weak-Form Sparse Additive DDA

The currently supported reliable Python estimator is polynomial WeakFormDDA. SparseAdditiveWeakFormDDA is experimental infrastructure for nonlinear main-effect delay discovery; use it for method development and diagnostics, not as a replacement for polynomial weak-form DDA in production analyses.

Fixed polynomial dictionaries are interpretable but can be brittle when the true delayed response is saturating, threshold-like, or otherwise smooth but not well approximated by low-order monomials. SparseAdditiveWeakFormDDA replaces polynomial terms with cubic B-spline response functions over candidate delays:

dx/dt = c + sum_tau f_tau(x(t - tau))

Each response function is fit in weak form by integrating spline basis functions over the same interval used for the observed increment. The first implementation is scalar and main-effect only; it does not yet fit pairwise delay surfaces. Delay selection is statistically meaningful only with group sparsity, because all spline basis functions for one delay must enter or leave the model together. Neighboring candidate delays can also be nearly duplicate predictors, so delay dictionary thinning and state-coverage diagnostics are part of the estimator.

from ddalab_app.backend.dda import (
    SparseAdditiveWeakFormDDA,
    stability_select_sparse_additive_weak_form,
)

model = SparseAdditiveWeakFormDDA(
    candidate_delays=[0, 10, 11, 24, 25, 40],
    n_knots=7,
    spline_degree=3,
    integration_window=9,
    stride=1,
    quadrature="trapezoid",
    regression="group_lasso",
    group_lasso_alpha=0.004,
    ridge_alpha=1e-6,
    standardize_features=True,
    thin_delays=True,
    max_delay_correlation=0.98,
)
model.fit([x_trial_1, x_trial_2, x_trial_3], dt=0.01)
print(model.selected_delays_)
print(model.group_strengths_)
print(model.delay_block_correlation_)
print(model.state_coverage_)

grid, effect = model.effect_curves_[10]

fit accepts either one scalar trajectory or a list of scalar trajectories. For list input, spline bases are shared across trajectories and weak-form windows are built within each trajectory only, never across trajectory boundaries.

Delay ranking is based on the L2 norm of each delay-specific spline coefficient block. With regression="group_lasso", inactive delay blocks can be driven to zero directly. For more conservative delay selection, use stability selection with null calibration:

scores = stability_select_sparse_additive_weak_form(
    [x_trial_1, x_trial_2, x_trial_3],
    dt=0.01,
    candidate_delays=model.candidate_delays_,
    integration_window=9,
    regression="group_lasso",
    group_lasso_alpha=0.004,
    n_repeats=20,
    null_repeats=10,
    calibrate=True,
)
print(scores.selection_frequency)
print(scores.null_selection_frequency)
print(scores.calibrated_selected_delays)

Prefer polynomial weak-form DDA when a low-order dictionary is scientifically meaningful and coefficient-level interpretation is the primary output. Prefer sparse additive weak-form DDA only when the main question is whether a small set of candidate delays has smooth nonlinear influence and the shape of that influence is itself informative. Current limitations are scalar input, main effects only, sensitivity to candidate-delay collinearity, dependence on broad state coverage, and validation-dependent choices of spline, group penalty, and stability thresholds.

Choosing Between WeakFormDDA and SparseAdditiveWeakFormDDA

Use WeakFormDDA as the default estimator when the goal is robust recovery of polynomial DDA coefficients or when the analysis must be stable without additional model-selection diagnostics. It is the reliable general-purpose Python estimator.

Use SparseAdditiveWeakFormDDA only as an experimental nonlinear-delay estimator. A defensible sparse-additive analysis should use group-lasso delay blocks, candidate-delay thinning, stability selection, null calibration, and state-coverage checks. Sparse-additive delay selections should not be trusted when instantaneous null diagnostics select delayed terms, even if interval prediction error is competitive. The current estimator also cannot represent pairwise or interacting nonlinear delayed effects; those require a future pairwise-surface model rather than the current main-effect spline model.

Desktop Development

cd packages/ddalab
./start.sh

The script provisions a local virtual environment, installs the unified package in editable mode, and launches the Qt application.

The desktop application uses a Qt Quick/QML workbench. Waveform and DDA-result plots use viewport-aware, cached Qt Quick scene-graph surfaces; there is no separate QWidget rendering mode.

See ARCHITECTURE.md for the desktop boundaries and data flow.

If you are working from source, ./start.sh expects cargo to be available so it can build or refresh the bundled dda-rs runtime.

Smoke Test

cd packages/ddalab
./start.sh --smoke-test