Errors in bar series are traditionally represented using error bars. However these give a poor visual fix on the data. By fading the bar rather than representing an error by a fixed-length bar, we better represent the uncertainty in the data. In the example, 'cat' and 'fox' catch approximately the same number of rabbits, whilst 'dog' catches more. This is more obvious in the faded representation.
The data is represented by a bar that fades out over the error, using a Gaussian to give the density of pixels. This gives a ragged upper surface, so the eye is not drawn to small differences in mean value.