Where the Middle Sits

There are three ways to say what is typical. The average: pour everyone's counters into one pile, then share them out evenly. The middle: line the numbers up and point to the one in the center. The most common: the number that shows up the most times.

Usually these three land close together. But one giant number can pull the average up while the middle barely moves — the average feels one huge value, and the middle does not.

The mean adds all values and divides by the count — the fair-share figure. The median is the middle value once the data are sorted; with an even count, average the two middle values. The mode is simply the value that occurs most often, and it is the only center that works on categories.

Because the mean uses every value's exact size, it is sensitive: a single very large or very small value shifts it. The median uses only position, so it hardly notices one extreme value. That difference is the heart of the next section.

The three centers answer different questions. The mean, x̄ = (Σx)/n, is the balance point of the data — the value about which the deviations sum to zero. The median is the 50th percentile, the value with half the data on each side. The mode is the most frequent value, the peak of the distribution. On a symmetric single-peaked shape all three coincide; asymmetry pulls them apart in a predictable order.

The mean is not resistant: because it weighs every value by its actual magnitude, a lone outlier or a long tail drags it toward the extreme. The median depends only on rank, so it is resistant — an outlier can grow without bound and the median stays put. This is why a right-skewed variable like income is almost always reported by its median: the mean would describe a household that scarcely exists.

Five friends line up by height. The one right in the middle shows the middle height. That is one way to say typical.

Another way: share all the sweets equally so everyone has the same number. That equal share is the average. And the flavour the most friends picked is the most common one.

Each center minimises a different loss. The mean minimises squared error, argmin Σ(x − c)², which is why it is the L² center and why it inherits the fragility of squared distance. The median minimises absolute error, argmin Σ|x − c|, an L¹ center whose optimality is flat across the middle, giving its celebrated robustness. The mode maximises density. The choice of center is, quietly, a choice of loss function.

Robustness has a precise currency: the breakdown point, the fraction of the data you may corrupt before the estimate can be driven to infinity. For the mean it is 0 — one bad value suffices. For the median it is 1/2, the highest possible. For a unimodal continuous distribution the empirical ordering mode, median, mean tracks the direction of skew, a rule of thumb (Pearson) that fails for some engineered distributions but holds across most real data.

As functionals on a distribution F, the mean is ∫x dF, the median is F⁻¹(1/2), and the mode is argmax of the density. Their influence functions expose the robustness gap: the mean's is unbounded and linear in x, so a single far observation exerts arbitrary leverage; the median's is bounded, a sign function scaled by 1/(2f(m)), so influence saturates. The asymptotic variance of the median, 1/(4 n f(m)²), depends on the density at the center rather than on the tails.

This trade is not free. At the normal model the median's efficiency relative to the mean is 2/π ≈ 0.64 — a third more variance, the price of robustness. The craft is matching estimator to contamination: under heavy tails a median or trimmed mean wins on mean squared error; under a clean Gaussian the mean is optimal. Reporting a center assumes a model of how the data can fail.