Notes
🔍

Geometric Mean

The geometric mean is a way to describe a “typical” value when numbers combine by multiplying rather than adding. It’s especially natural for things like growth rates, ratios, and any situation where changes compound over time.

What it is (and how it’s computed)
The geometric mean of positive numbers is the nth root of their product. For values \(x_1, x_2, \dots, x_n\), it is: 

GM = (x1 * x2 * ... * xn)^(1/n)

A common equivalent form uses logs (often more numerically stable): take the average of the logs, then exponentiate.

Why it differs from the arithmetic mean
The arithmetic mean treats changes as additive (“+10, +10”). The geometric mean treats them as multiplicative (“×1.10, ×1.10”). That makes it the right “center” for:

  • Compounded growth (returns, population growth, inflation)
  • Ratios and rates (speedups, fold-changes in biology)
  • Skewed positive data where large values shouldn’t dominate as much

Practical examples

  • Investment returns: If a stock goes +20% then −20%, the arithmetic average return is 0%, but the geometric mean return is \(\sqrt{1.2 \times 0.8}-1 \approx -2.0\%\), reflecting the real compounding loss.
  • Model evaluation: In information retrieval and classification, metrics like F1 relate to the geometric mean idea (balancing multiplicative trade-offs between precision and recall).

Why it matters in AI/ML
In ML workflows, the geometric mean shows up when averaging multiplicative factors: combining likelihoods, summarizing perplexity-like quantities, or aggregating ratios across folds. It’s also tied to the log transform: averaging logs (additive) corresponds to a geometric mean in the original scale, which is a common trick for stabilizing variance and handling skewed targets.

Geometric Mean is the nth root of the product of n positive values, equivalently exp(average of their logs). It summarizes multiplicative processes and relative changes, making it preferable to the arithmetic mean for ratios, growth rates, and log-normal data. In AI/ML, it is used to average fold-wise performance ratios or learning-rate multipliers. Example: average return over days is the geometric mean of (1+r) factors, minus 1.

Imagine you’re tracking how your money grows: +10% one year, then +20% the next. You wouldn’t just “average” 10 and 20, because growth stacks on top of growth. The geometric mean is the kind of average that fits situations where values multiply together, like growth rates, ratios, or repeated percentage changes.

In statistics (and in AI/ML), the geometric mean gives a “typical” value for multiplicative data, helping summarize things like average model improvement factors or average change across many steps. It’s especially useful when numbers vary a lot and you care about proportional change.