To visualize uncertainty of measurement in our dataset or to indicate the error, we can use error bars. Error bars can easily give an idea of how error free the dataset is. They can show one standard deviation, one standard error, or 95 percent confidence interval. There is no standard here, so always explicitly state what values (errors) error bars display. Most papers in the experimental sciences should contain error bars to present accuracy of the data.

Even though just two parameters are mandatory—left and height— we often want to use more than that. Here are some parameters we can use:

Some optional arguments (color, edgecolor, linewidth, xerr, and yerr) can be single values or sequences with the same length as the number of bars.

Let's illustrate this using an example:

import numpy as np
import matplotlib.pyplot as plt

# generate number of measurements
x = np.arange(0, 10, 1)

# values computed from "measured"
y = np.log(x)

# add some error samples from standard normal distribution
xe = 0.1 * np.abs(np.random.randn(len(y)))

# draw and show errorbar
plt.bar(x, y, yerr=xe, width=0.4, align='center', ecolor='r', color='cyan', label='experiment #1');

# give some explanations
plt.xlabel('# measurement')
plt.ylabel('Measured values')
plt.title('Measurements')
plt.legend(loc='upper left')

plt.show()

The preceding code will plot the following diagram:

To be able to plot an error bar, we needed to have some measures (x); for every measure computed (y), we introduced errors (xe).

We used NumPy to generate and compute values; standard distributions are good enough for demonstration purposes, but if you happen to know your data distribution in advance, you can always make some prototype visualizations and try out different layouts to find the best options to present information.

Another interesting option to use if we are preparing visualizations for a black-and-white medium is hatch; it can have the following values:

What we have just used are error bars known as symmetrical error bars. If the nature of our dataset is such that errors are not the same in both directions (negative and positive), we can also specify them separately using asymmetrical error bars.

All we have to do differently is to specify xerr or yerr using a two-element list (such as a 2D array), where the first list contains values for negative errors and the second one for positive errors.