Configure multistart

LinearResponse creates n_init_cond random initial displacement and velocity vectors for every requested frequency. Its sampling bounds are based on modal linear-response scales:

\[ x_{\max,i} = \frac{Q_i}{\omega_{0,i}^2}c, \qquad v_{\max,i} = \frac{Q_i}{\omega_{0,i}}c, \]

where c is linear_response_factor. The strategy samples uniformly inside the resulting hyper-rectangle using random_seed for reproducibility.

multistart = poscidyn.LinearResponse(
    n_init_cond=32,
    linear_response_factor=1.5,
    random_seed=0,
)

Increase n_init_cond when coexisting responses are likely. Increase linear_response_factor when the expected nonlinear amplitude lies outside the linear scale. Both increase the computational search burden, and a larger batch also increases memory use.

These are heuristics for exploring basins of attraction, not a branch-complete sampling guarantee. Validate sensitive results with changed seeds and denser sampling.