ScoreQuant

ScoreQuant

ScoreQuant sorts events into a few hard bins, chosen so that the bin counts keep as much information about your model parameters as possible, and it reports how much they kept.

pip install scorequant Get started →

Three panels. Top left: the raw score plane tinted by six convex cells with straight edges, the trajectory coloured by cell. Top right: the same cells after whitening, an ordinary Voronoi diagram of the cell means. Bottom: the same labels along the detector, a comb of many intervals above six equal segments.
Bins for an interferometer, from the Michelson walkthrough. In score space (left) each bin is a convex cell. Along the detector (bottom) the same bins form a comb, very unlike equal segments.

Why score space

Near a reference point, all an estimate uses from an event is its score s(x)=∇θlog⁡p(x∣θ)s(x)=\nabla_\theta \log p(x\mid\theta), one number per parameter. A bin keeps only its total weight and the mean score of its events, so the information it loses is the spread of scores inside it. Events with similar scores should therefore share a bin, however far apart they are in the original observation, and neighbours with different scores should not. ScoreQuant draws the bins in score space to keep information in every parameter direction at once (D-optimality).

Walkthroughs

  • Phase Estimation in a Michelson Interferometer

    Equal detector segments can lose the phase entirely; cells drawn in score space keep it, and the D-optimal cells compile into a reusable rule.

  • From a classifier to Fisher-preserving bins

    A calibrated classifier can stand in for the likelihood, but the retention ScoreQuant then reports describes the classifier's estimate, not the data.

  • A Higgs boson measurement

    Once the energy scale floats, cuts on the classifier output lose much of their precision on the signal strength; a rule fitted with the energy scale profiled out loses little.

  • Bone-marrow cell populations

    Reduced to 8 counts each, held-out patients give population fractions with an error of the same order as using every cell.

The benchmark sets each result beside equal-width, quantile and k-means bins.

Get started

Get started fits a reusable rule on a small example, applies it to new data and saves it. The user guide covers inputs, criteria, solvers and nuisance parameters. The research page states the theorem behind the rule and its limits.