A normalized histogram is a histogram where the raw frequency counts have been rescaled so the values represent proportions or densities instead of raw counts. There are two common ways to normalize, and they produce different y-axis values, so understanding which one you’re looking at matters.
Two Types of Normalization
A regular histogram counts how many data points fall into each bin and plots those counts on the y-axis. That’s useful on its own, but the numbers are tied to your specific sample size. Normalization removes that dependency. The two approaches differ in what they guarantee about the result.
Relative frequency normalization: Divide the count in each bin by the total number of observations. The height of each bar now represents the proportion of data in that bin, and all the bar heights sum to 1 (or 100%). This is the simpler version, and the y-axis values are always between 0 and 1.
Probability density normalization: Divide the count in each bin by both the total number of observations and the bin width. Now it’s the total area of all bars that equals 1, not the sum of the heights. This is the version most people mean when they say “normalized histogram” in a statistics or data science context.
Why Bin Width Matters
The density version exists because of a subtlety: bin width affects how a histogram looks. If you use narrow bins, bar heights drop. If you use wide bins, bar heights rise. Dividing by bin width cancels out that effect, giving you a quantity (density) that stays comparable regardless of how you chose to slice up the x-axis.
The formula for each bar’s height in a density-normalized histogram is:
height = count in bin / (total observations × bin width)
Because you’re dividing by the bin width, each bar’s area (height × width) equals the proportion of data in that bin. Add up all those areas and you get exactly 1. This is the same property that defines a probability density function in statistics, which is why this type of histogram is so useful for estimating distributions.
Why Bar Heights Can Exceed 1
A common point of confusion: in a density-normalized histogram, individual bar heights can be greater than 1. That’s perfectly normal. The constraint is on the total area, not on any single bar’s height. If your data is tightly clustered and your bins are narrow, a bar might reach 2, 5, or higher. The height represents density, not probability. To get the actual probability of a value falling in a particular bin, you multiply that bar’s height by the bin width, which always gives a number between 0 and 1.
Comparing Datasets With Different Sample Sizes
The main practical reason to normalize a histogram is comparison. Suppose you have one dataset with 500 observations and another with 10,000. Plotting raw counts side by side is misleading because the larger dataset will dominate visually. Its bars will be much taller even if the two datasets follow the same distribution.
Scaling both histograms to area 1 puts them on equal footing. The shapes become directly comparable because you’re looking at density rather than volume of data. This is the standard approach when you want to visually assess whether two samples come from similar distributions.
How Normalized Histograms Relate to Probability Distributions
A density-normalized histogram is an empirical estimate of the underlying probability density function (PDF) of your data. The PDF is the smooth theoretical curve that describes how likely different values are for a continuous variable. Its total area under the curve equals 1, just like the total area of a density histogram.
As you collect more data and use more bins, the jagged shape of the histogram gradually approaches the smooth PDF curve. This is why you’ll often see a smooth curve overlaid on a normalized histogram: the histogram shows what your data actually looks like, and the curve shows the theoretical distribution you’re comparing it to. For that overlay to make sense, both need to be on the same scale, which means the histogram needs density normalization.
How Software Handles It
Most plotting libraries have a built-in option for density normalization. In Python’s Matplotlib, setting density=True in the histogram function divides each bin’s count by the total count multiplied by the bin width. The result integrates to 1 across the full range of data. The internal calculation is straightforward: density equals counts divided by the sum of all counts times the bin widths.
In R, the hist() function accepts a freq=FALSE argument that produces the same density scaling. Other tools like NumPy, MATLAB, and Excel can do it too, though the parameter names vary. If your software doesn’t have a built-in option, you can normalize manually by applying the formula: take each bin’s count, divide by the total number of data points, then divide again by the bin width.
Choosing the Right Normalization
Use relative frequency normalization (bar heights sum to 1) when you want to quickly see what percentage of your data falls in each bin. This is intuitive and works well for presentations where your audience isn’t used to reading density plots.
Use density normalization (total area equals 1) when you want to overlay a theoretical distribution, compare datasets with different sample sizes, or when your bins aren’t all the same width. Unequal bin widths make relative frequency histograms visually misleading because wider bins look disproportionately important. Density normalization corrects for this automatically since the division by bin width accounts for the difference.
If all your bins are the same width and you’re not comparing across datasets or fitting a distribution, a raw count histogram is often the clearest choice. Normalization adds value when you need the histogram to represent something beyond “how many data points landed here.”