NumPy Histogram tricks for HEP

2018-02-08

Update August 2019: About a year after writing this blog post I created a Python package to handle all of my pythonic histogramming needs. It's called pygram11. This post is definitely still useful for learning more details about NumPy histogramming.

Our starting point

Histogramming some data is simple using numpy.histogram.

>>> import numpy as np
>>> x = np.random.randn(10000)           ## create a dataset
>>> w = np.random.normal(1, 0.2, 10000)  ## create some phony weights
>>> b = np.linspace(-5, 5, 11)           ## bin edges (10 bins from -5 to 5)
>>> n, bins = np.histogram(x, bins=b, weights=w)

This gives me two arrays

Quick and simple -- but what if I want to include underflow and overflow in the first and last bins, respectively? What if I want to compute the error on each bin height given a weighted dataset? These quantities are important for high energy physics, where nearly all of our analysis is done using histograms.

Underflow and overflow

Where the elements of the data contribute to the bin height is of course determined by the bin edges. We can make the left and right edges infinite to be sure to include all of our data1. Then we just add the [0] bin contents to the [1] bin contents, and add the [-1] bin contents to the [-2] bin contents. Finally, we polish it off by chopping off the out-of-bounds elements:

>>> import numpy as np
>>> raw_bins = np.linspace(-5, 5, 11)
>>> use_bins = [np.array([-np.inf]), raw_bins, np.array([np.inf])]
>>> use_bins = np.concatenate(use_bins)
>>> x = np.random.normal(0, 2, 1000) ## phony dataset
>>> n, bins = np.histogram(x, bins=use_bins)
>>> n[1]  += n[0]   ## add underflow to first bin
>>> n[-2] += n[-1]  ## add overflow to last bin
>>> n = n[1:-1]     ## chop off the under/overflow
>>> bins = raw_bins ## use our original binning (without infinities)

And that's it, now all of the data is histogrammed -- including under and overflow.

Error on bin height using weights

The standard error on a bin height is simply the square-root of the bin height, \(\sqrt{N}\)2. If a bin is constructed from weighted data, we require the square-root of the sum of the weights squared, \(\sqrt{\sum_i w_i^2}\).

The numpy.histogram function doesn't provide any information about which weights belong to which bin, but we have another useful NumPy function which can generate an array of indices based on where data falls in a particular set of bins, numpy.digitize.

First, we get an array representing which bin each data point would fall into. We can then use the conditional function numpy.where in a loop over all bins to grab only the weights in that bin, and sum their squares.

>>> import numpy as np
>>> x = np.random.normal(0, 2.0, 1000)         ## a dataset
>>> b = np.linspace(-2, 2, 21)                 ## 20 bins
>>> w = np.random.normal(1, 0.2, 1000)         ## some weights
>>> sum_w2 = np.zeros([20], dtype=np.float32)  ## start with empty errors
>>> digits = np.digitize(x, b)                 ## bin index array for each data element
>>> for i in range(nbins):
>>>     weights_in_current_bin = w[np.where(digits == i)[0]]
>>>     sum_w2[i] = np.sum(np.power(weights_in_current_bin, 2))
>>> n, bins = np.histogram(x, bins=b, weights=w)
>>> err = np.sqrt(sum_w2)

Now two arrays exist: n contains the heights in each bin, and err contains the standard error on the bin heights.

Appendix, a function to combine the two methods:

def extended_hist(
    x: np.ndarray
    nbins: int
    range: Tuple[float, float],
    underflow: bool = True,
    overflow: bool = True,
    weights: Optional[np.ndarray] = None,
) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
    """Histogram weighted data with potential under/overflow.

    Parameters
    ----------
    x : array_like
        Data to histogram.
    nbins : int
        Total number of bins.
    range : (float, float)
        Definition of binning max and min.
    underflow : bool
        Include undeflow data in the first bin.
    overflow : bool
        Include overflow data in the last bin.
    weights : array_like, optional
        Weights associated with each element of ``x``.

    Returns
    -------
    numpy.ndarray
        Total bin values.
    numpy.ndarray
        Poisson uncertainty on each bin count.
    numpy.ndarray
        Bin centers.
    numpy.ndarray
        Bin edges.

    """
    if weights is not None:
        if weights.shape != x.shape:
            raise ValueError(
                "Unequal shapes x: {}; weights: {}".format(
                    x.shape, weights.shape
                )
            )
    xmin, xmax = range
    edges = np.linspace(xmin, xmax, nbins + 1)
    neginf = np.array([-np.inf], dtype=np.float32)
    posinf = np.array([np.inf], dtype=np.float32)
    bins = np.concatenate([neginf, edges, posinf])
    if weights is None:
        hist, bin_edges = np.histogram(x, bins=bins)
    else:
        hist, bin_edges = np.histogram(x, bins=bins, weights=weights)

    n = hist[1:-1]
    if underflow:
        n[0] += hist[0]
    if overflow:
        n[-1] += hist[-1]

    if weights is None:
        u = np.sqrt(n)
    else:
        bin_sumw2 = np.zeros(nbins + 2, dtype=np.float32)
        digits = np.digitize(x, edges)
        for i in range(nbins + 2):
            bin_sumw2[i] = np.sum(
                np.power(weights[np.where(digits == i)[0]], 2)
            )
        u = bin_sumw2[1:-1]
        if underflow:
            u[0] += bin_sumw2[0]
        if overflow:
            u[-1] += bin_sumw2[-1]
        u = np.sqrt(u)

    centers = np.delete(edges, [0]) - (np.ediff1d(edges) / 2.0)
    return n, u, centers, edges

Update August 2019: With pygram11, we can just import the histogram function and call a one-liner for the values and the error:

>>> from pygram11 import histogram
>>> data, weights = get_some_weighted_data()
>>> h, err = histogram(data, bins=10, range=(xmin, xmax), weights=weights, flow=True)

1

In the implementation of numpy.histogram, elements of the input array that live outside the bounds of the binning are ignored.

2

Bin height is related to counting, therefore the data in a bin is Poissonian. The variance of a Poisson distribution is \(N\).