PiecewiseWLS

PiecewiseWLS

CLI: --model piecewise_wls. Class: loqculate.models.piecewise_wls.PiecewiseWLS.

Same mean function as PiecewiseCF: $y = \max(c, ax+b)$. Same default weights. Different solver.

I keep this class so I can compare TRF against the closed-form search on the same peptides. It is not the CLI default.

Solver

scipy.optimize.curve_fit with method defaulting to TRF, maxfev=5000, bounds $a \ge 0$, $c-b \ge 0$, $b$ unbounded. The curve is parameterized as $(a, b, c-b)$ so the noise-above-linear constraint is a bound.

init_method default is "legacy": slope from the mean signals at the two highest unique concentrations, noise intercept from the lowest, linear intercept from the top point. "auto" is a WLS line above the two lowest concentrations. The CLI does not expose init_method; a CLI piecewise_wls fit is legacy init.

If TRF raises RuntimeError or ValueError, fit() stores the clamped initial guess and does not raise.

If np.ptp(y) == 0 (constant area, including all zeros), TRF is skipped. Slope is 0 and both intercepts equal that constant. That path exists because a tiny positive TRF slope on a flat curve produced a finite LOD on arm64.

params_ has slope, intercept_linear, intercept_noise. There is no knot_x.

LOD and LOQ

LOD uses the same arithmetic as PiecewiseCF: intersection $(c-b)/a$, noise observations $x < x_{\mathrm{int}}$, ddof=1 std, $\mathrm{LOD}=(c+\mathrm{std_mult}\cdot\sigma-b)/a$, then the min_noise_points / min_linear_points / max(x) guards.

LOQ is bootstrap only. loq() has no method argument. Grid, window, and <= threshold match PiecewiseCF (grid_points=100, sliding_window=3, cv_thresh=0.20). The window is consecutive grid points.

There is no loq_delta and no CLI --fast on this model.

Constructor defaults

init_method="legacy", n_boot_reps=100, seed=42, min_noise_points=2, min_linear_points=1, sliding_window=3, grid_points=100. fit() requires at least 3 observations.