Binlets: denoising via adaptive binning

Usage

binlets requires an array of data, and a test function to compare two data points.

Single-channel with Poisson statistics

For instance, for single-channel signal with Poisson statistics:

def chi2_test(x: NDArray, y: NDArray) -> NDArray[bool]:
    """Compares two values with Poisson noise using a χ² test."""
    diff = x - y
    var_diff = x + y  # Poisson variance
    return diff**2 <= var_diff


denoised = binlets(
    data,
    test=chi2_test,
    levels=None,  # max averaging area is 2**levels. By default, floor(log2(min(data.shape)))
    linear=True,  # the transformation is linear (x - y)
)

We recomend wrapping this in a function, and providing an extra parameter to adjust the significance level:

def poisson_binlets(data: NDArray, *, n_sigma: float, levels: int | None = None):
    def chi2_test(x: NDArray, y: NDArray) -> NDArray[bool]:
        """Compares two values with Poisson noise using a χ² test."""
        diff = x - y
        var_diff = x + y  # Poisson variance
        return diff**2 <= n_sigma**2 * var_diff

    denoised = binlets(
        data,
        test=chi2_test,
        linear=True,
    )
    return denoised

Ratio of multichannel-data

For multichannel data, binlets expects channels to be in the first dimension of the data array. That is, data.shape should be (N_CHANNELS, *spatial_dimensions).

def ratio(channels):
    """Statistic of interest."""
    return channels[1] / channels[0]


def test(x, y):
    # The test of your choice. For instance, a χ² test.
    diff = ratio(x) - ratio(y)
    var_diff = ratio_var(x) + ratio_var(y)
    return diff**2 <= var_diff


denoised = binlets(data, test=test, ...)  # the same as before
denoised_ratio = ratio(denoised)

Installation

Binlets can be installed from PyPI:

pip install binlets

or conda-forge:

conda install -c conda-forge binlets

Development

We use pixi for development. Just clone the repository and run pixi run test. Optionally, install pre-commit hooks with pixi run prek install.