fit_metric()

FeatureMetric MayaFlux::Kinesis::fit_metric ( std::span< const Eigen::VectorXd >  samples, const Eigen::VectorXd &  periods = {}  )

inline

Derive per-axis metric weights from the spread of a sample set.

Parameters

samples	Observations, all of the same dimension
periods	Per-axis wrap periods, or an empty vector for none

Returns

A metric whose weights are the reciprocal of each axis's standard deviation, so one standard deviation on any axis contributes equally to distance

The answer to incommensurate units when a corpus is available: rather than the caller guessing that pressure should count twice as much as speed, the weights come from how much each axis actually varied across the observations. An axis with no variation gets unit weight rather than an infinite one.

Definition at line 240 of file FeatureExtent.hpp.

                                   {})
{
    if (samples.empty())
        return {};
 
    const Eigen::Index n = samples.front().size();
    Eigen::VectorXd mean = Eigen::VectorXd::Zero(n);
    for (const auto& s : samples)
        mean += s;
    mean /= static_cast<double>(samples.size());
 
    Eigen::VectorXd var = Eigen::VectorXd::Zero(n);
    for (const auto& s : samples) {
        const Eigen::VectorXd d = s - mean;
        var += d.cwiseProduct(d);
    }
    var /= static_cast<double>(samples.size());
 
    Eigen::VectorXd weights(n);
    for (Eigen::Index i = 0; i < n; ++i) {
        const double sd = std::sqrt(var(i));
        weights(i) = (sd > 1e-12) ? (1.0 / sd) : 1.0;
    }
 
    return { .weights = weights,
        .periods = (periods.size() == n) ? periods : Eigen::VectorXd::Zero(n) };
}