"""
Functions for least-squares fitting with optional regularization.
"""
from warnings import warn
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LinearRegression, Ridge
[docs]def least_squares(jacobian, data, weights, damping=None):
"""
Solve a weighted least-squares problem with optional damping regularization.
Scales the Jacobian matrix so that each column has unit variance. This helps keep
the regularization parameter in a sensible range. The scaling is undone before
returning the estimated parameters so that scaling isn't required for predictions.
Doesn't normalize the column means because that operation can't be undone.
Parameters
----------
jacobian : 2d-array
The Jacobian/sensitivity/feature matrix.
data : 1d-array
The data array. Must be a single 1D array. If fitting multiple data components,
stack the arrays and the Jacobian matrices.
weights : None or 1d-array
The data weights. Like the data, this must also be a 1D array. Stack the weights
in the same order as the data. Use ``weights=None`` to fit without weights.
damping : None or float
The positive damping (Tikhonov 0th order) regularization parameter. If
``damping=None``, will use a regular least-squares fit.
Returns
-------
parameters : 1d-array
The estimated 1D array of parameters that fit the data.
"""
if jacobian.shape[0] < jacobian.shape[1]:
warn(
"Under-determined problem detected (ndata, nparams)={}.".format(
jacobian.shape
)
)
scaler = StandardScaler(copy=False, with_mean=False, with_std=True)
jacobian = scaler.fit_transform(jacobian)
if damping is None:
regr = LinearRegression(fit_intercept=False, normalize=False)
else:
regr = Ridge(alpha=damping, fit_intercept=False, normalize=False)
regr.fit(jacobian, data.ravel(), sample_weight=weights)
params = regr.coef_ / scaler.scale_
return params