| .. _astropy-visualization-hist: |
|
|
| *********************** |
| Choosing Histogram Bins |
| *********************** |
|
|
| The :mod:`astropy.visualization` module provides the |
| :func:`~astropy.visualization.hist` function, which is a generalization of |
| matplotlib |
| of histogram bins. For computing bins without the accompanying plot, see |
| :func:`astropy.stats.histogram`. |
|
|
| As a motivation for this, consider the following two histograms, which are |
| constructed from the same underlying set of 5000 points, the first with |
| matplotlib |
| 200 bins: |
|
|
| .. plot:: |
| :align: center |
|
|
| import numpy as np |
| import matplotlib.pyplot as plt |
|
|
| # generate some complicated data |
| rng = np.random.RandomState(0) |
| t = np.concatenate([-5 + 1.8 * rng.standard_cauchy(500), |
| -4 + 0.8 * rng.standard_cauchy(2000), |
| -1 + 0.3 * rng.standard_cauchy(500), |
| 2 + 0.8 * rng.standard_cauchy(1000), |
| 4 + 1.5 * rng.standard_cauchy(1000)]) |
|
|
| # truncate to a reasonable range |
| t = t[(t > -15) & (t < 15)] |
|
|
| # draw histograms with two different bin widths |
| fig, ax = plt.subplots(1, 2, figsize=(10, 4)) |
|
|
| fig.subplots_adjust(left=0.1, right=0.95, bottom=0.15) |
| for i, bins in enumerate([10, 200]): |
| ax[i].hist(t, bins=bins, histtype= |
| ax[i].set_xlabel( |
| ax[i].set_ylabel( |
| ax[i].set_title( |
| fontdict=dict(family= |
|
|
| Upon visual inspection, it is clear that each of these choices is suboptimal: |
| with 10 bins, the fine structure of the data distribution is lost, while with |
| 200 bins, heights of individual bins are affected by sampling error. |
| The tried-and-true method employed by most scientists is a trial and error |
| approach that attempts to find a suitable midpoint between these. |
|
|
| Astropy |
| providing several methods of automatically tuning the histogram bin size. |
| It has a syntax identical to matplotlib |
| exception of the ``bins`` parameter, which allows specification of one of |
| four different methods for automatic bin selection. These methods are |
| implemented in :func:`astropy.stats.histogram`, which has a similar syntax |
| to the ``np.histogram`` function. |
|
|
| Normal Reference Rules |
| ====================== |
| The simplest methods of tuning the number of bins are the normal reference |
| rules due to Scott (implemented in :func:`~astropy.stats.scott_bin_width`) and |
| Freedman & Diaconis (implemented in :func:`~astropy.stats.freedman_bin_width`). |
| These rules proceed by assuming the data is close to normally-distributed, and |
| applying a rule-of-thumb intended to minimize the difference between the |
| histogram and the underlying distribution of data. |
|
|
| The following figure shows the results of these two rules on the above dataset: |
|
|
| .. plot:: |
| :align: center |
|
|
| import numpy as np |
| import matplotlib.pyplot as plt |
| from astropy.visualization import hist |
|
|
| # generate some complicated data |
| rng = np.random.RandomState(0) |
| t = np.concatenate([-5 + 1.8 * rng.standard_cauchy(500), |
| -4 + 0.8 * rng.standard_cauchy(2000), |
| -1 + 0.3 * rng.standard_cauchy(500), |
| 2 + 0.8 * rng.standard_cauchy(1000), |
| 4 + 1.5 * rng.standard_cauchy(1000)]) |
|
|
| # truncate to a reasonable range |
| t = t[(t > -15) & (t < 15)] |
|
|
| # draw histograms with two different bin widths |
| fig, ax = plt.subplots(1, 2, figsize=(10, 4)) |
|
|
| fig.subplots_adjust(left=0.1, right=0.95, bottom=0.15) |
| for i, bins in enumerate([ |
| hist(t, bins=bins, ax=ax[i], histtype= |
| alpha=0.2, density=True) |
| ax[i].set_xlabel( |
| ax[i].set_ylabel( |
| ax[i].set_title( |
| fontdict=dict(family= |
|
|
|
|
| As we can see, both of these rules of thumb choose an intermediate number of |
| bins which provide a good trade-off between data representation and noise |
| suppression. |
|
|
| Bayesian Models |
| =============== |
|
|
| Though rules-of-thumb like Scott |
| fast and convenient, their strong assumptions about the data make them |
| suboptimal for more complicated distributions. Other methods of bin selection |
| use fitness functions computed on the actual data to choose an optimal binning. |
| Astropy implements two of these examples: Knuth |
| :func:`~astropy.stats.knuth_bin_width`) and Bayesian Blocks (implemented in |
| :func:`~astropy.stats.bayesian_blocks`). |
|
|
| Knuth |
| histogram |
| flexible method which allows varying bin widths. Because both of these require |
| the minimization of a cost function across the dataset, they are more |
| computationally intensive than the rules-of-thumb mentioned above. Here are |
| the results of these procedures for the above dataset: |
|
|
| .. plot:: |
| :align: center |
|
|
| import warnings |
| import numpy as np |
| import matplotlib.pyplot as plt |
| from astropy.visualization import hist |
|
|
| # generate some complicated data |
| rng = np.random.RandomState(0) |
| t = np.concatenate([-5 + 1.8 * rng.standard_cauchy(500), |
| -4 + 0.8 * rng.standard_cauchy(2000), |
| -1 + 0.3 * rng.standard_cauchy(500), |
| 2 + 0.8 * rng.standard_cauchy(1000), |
| 4 + 1.5 * rng.standard_cauchy(1000)]) |
|
|
| # truncate to a reasonable range |
| t = t[(t > -15) & (t < 15)] |
|
|
| # draw histograms with two different bin widths |
| fig, ax = plt.subplots(1, 2, figsize=(10, 4)) |
|
|
| fig.subplots_adjust(left=0.1, right=0.95, bottom=0.15) |
| for i, bins in enumerate([ |
| with warnings.catch_warnings(): |
| warnings.simplefilter( |
| hist(t, bins=bins, ax=ax[i], histtype= |
| alpha=0.2, density=True) |
| ax[i].set_xlabel( |
| ax[i].set_ylabel( |
| ax[i].set_title( |
| fontdict=dict(family= |
|
|
|
|
| Notice that both of these capture the shape of the distribution very |
| accurately, and that the ``bins= |
| in width depending on the local structure in the data. Compared to standard |
| defaults, these Bayesian optimization methods provide a much more principled |
| means of choosing histogram binning. |
|
|