Loss Functions

Loss functions control how residuals between data and model are computed during fitting. Pass a loss function as the loss argument to any model's fit() method.

chi2_loss

danish.loss.chi2_loss module-attribute

chi2_loss = systematic_loss(0)

Standard chi residual, equivalent to systematic_loss(0).

Assumes Gaussian read noise plus Poisson shot noise: (data - model) / sqrt(var + model). This is the default loss function.

systematic_loss

danish.loss.systematic_loss

systematic_loss(alpha)

Return a loss function that adds a fractional systematic noise floor.

The effective variance is var + model + (alpha * model)**2, where alpha is the fractional systematic uncertainty (e.g. 0.02 = 2%). This caps the effective SNR per pixel, preventing unmodeled correlated residuals (e.g. from atmospheric fluctuations or mirror figure errors) from dominating the fit.

Parameters:

Name	Type	Description	Default
alpha	float	Fractional systematic uncertainty. Typical values are 0.01–0.05.	required

Returns:

Type	Description
callable	Loss function with signature (data, model, var) -> array.

Source code in danish/loss.py

def systematic_loss(alpha):
    """Return a loss function that adds a fractional systematic noise floor.

    The effective variance is ``var + model + (alpha * model)**2``, where
    ``alpha`` is the fractional systematic uncertainty (e.g. 0.02 = 2%).
    This caps the effective SNR per pixel, preventing unmodeled correlated
    residuals (e.g. from atmospheric fluctuations or mirror figure errors)
    from dominating the fit.

    Parameters
    ----------
    alpha : float
        Fractional systematic uncertainty.  Typical values are 0.01–0.05.

    Returns
    -------
    callable
        Loss function with signature ``(data, model, var) -> array``.
    """
    def _loss(data, model, var):
        return (data - model) / np.sqrt(var + model + (alpha * model)**2)
    return _loss