The reasonable effectiveness of Fermi estimates

January 4, 2021. Why are Fermi approximations so effective? One important factor is log normality, which occurs for large random products, also related to the mechanism underlying the Newcomb-Benford law for first digits. Another element is variance-reduction through subestimates, and I offer some tips on how to subestimate judiciously.


Fermi approximation is the art of making good order-of-magnitude estimates. I’ve written about them at greater length here and here, but I’ve never really found a satisfactory explanation for why they work. Order-of-magnitude is certainly a charitable margin of error, but time and time again, I find they are better than they have any right to be! Clearly, there must be an underlying statistical explanation for this unreasonable effectiveness.

There are two key techniques: the use of geometric means, and the factorisation into subestimates. We will try to explain the first using logarithmic uniformity, which is the same mechanism underlying the anomalous distribution of first digits known as the Newcomb-Benford law. We give a looser but related explanation of the second in terms of strategies for variance-reduction in human error.

Products and logarithmic uniformity

Suppose a random variable $F$ is a product of many independent random variables,

\[F = X_1 X_2 \cdots X_N.\]

Then the logarithm of $F$ is a sum of many random variables $Y_i = \log X_i$:

\[\log F = \log X_1 + \log X_2 + \cdots + \log X_N = \sum_{i=1}^N Y_i.\]

By the central limit theorem for unlike variables (see e.g. this post), for large $N$ this approaches a normal distribution:

\[\log F \to \mathcal{N}(\mu, \sigma^2), \quad \mu := \sum_i \mu_i, \quad \sigma^2 = \sum_i \sigma_i^2,\]

where the $Y_i$ has mean $\mu_i$ and variance $\sigma_i^2$. We say that $F$ has a log-normal distribution, since its log is normal. To get uniformity into the picture, we can zoom in on the region near $F = e^\mu$ where the probability density is approximately uniform. More carefully, the density is

\[p(x) = \frac{1}{\sigma\sqrt{2\pi}} e^{-(x-\mu)^2/2\sigma^2}.\]

Taylor-expanding near $x = \mu$ gives

\[p(x) = \frac{1}{\sigma\sqrt{2\pi}} \left[1 - \frac{(x-\mu)^2}{2\sigma^2} + O(x^4)\right].\]

This looks uniform provided $(x - \mu)^2 \ll \sigma^2$. For instance, at a third of a standard deviation, $x = \mu + \sigma/3$, we have

\[1 - \frac{(x-\mu)^2}{2\sigma^2} = 1 - \frac{1}{18} \approx 0.94,\]

and $\text{erf}(1/\sqrt{18}) \approx 0.26$, about a quarter of the probability mass, lies underneath. This is what we mean when we say that $F$ is logarithmically uniform.

Geometric means

In Fermi estimates, one of the basic techniques is to take geometric means of estimates, typically an overestimate and an underestimate. For instance, to Fermi estimate the population of Chile, I could consider a number like one million which seems much too low, and a number like one hundred million which seems much too high, and take their geometric mean:

\[\sqrt{(1 \text{ million}) \times (100 \text{ million})} = 10 \text{ million}.\]

Since population is a product of many different factors, it is reasonable to expect it to approximate a log-normal distribution. Then, after logs, the geometric mean $\sqrt{ab}$ becomes the arithmetic mean of $\log a$ and $\log b$:

\[\log \sqrt{ab} = \frac{1}{2}(\log a + \log b).\]

Taking the mean $\mu$ of the distribution as the true value, these geometric means provide an unbiased estimator of the mean. Moreover, the error of the estimate will decrease as $1/k$ for $k$ samples (assuming human estimates sample from the distribution), so more is better. To see how much better I could do on the Chile population estimate, I solicited guesses from four friends, and obtained $20, 20, 30$ and $35$ million. Combining with my estimate, I get a geometric mean

\[(10 \times 20 \times 20 \times 30 \times 35)^{1/5} \text{ million} \approx 21 \text{ million}.\]

The actual population is around $18$ million, so the estimate made from more guesses is indeed better! This is also better than the arithemetic average, $23$ million. Incidentally, this also illustrates the wisdom of the crowd, also called “diversity of prediction”. The individual errors from a broad spread of guesses tend to cancel each other out, leading to a better-behaved average, though in this case in logarithmic space.

In general, Fermi estimates work best for numbers which are large random products (this is how we try to solve them!), so the problem domain tends to enforce the statistical properties we want. For many examples of log-normal distributions in the real world, see Limpert, Stahel and Abbt (2001). It’s worth noting that not everything we can Fermi estimate is log-normal, however. Many things in the real world obey power laws, for instance, and although you can exploit this to make better Fermi estimates (as lukeprog does in his tutorial), we can happily Fermi estimate power-law distributed numbers without this advanced technology. Are Fermi estimates unreasonably effective in this context? Maybe. But the estimates work best in the high-density core where things look uniform, not out at the tails, and it’s not until we get to the tails that the difference between the log-normal and power law (or exponential, or Weibull, or your favourite skewed distribution) becomes pronounced. So the unreasonable effectiveness here can probably be explained by the resemblance to the log-normal, though this is something I’d like to check more carefully in future.

The Newcomb-Benford law

Logarithmic uniformity also explains an odd pattern in the first digits of naturally occurring numbers like tax returns, stock market prices, populations, river lengths, physical constants, and even powers of $2$. The pattern, called the Newcomb-Benford law after Simon Newcomb and Frank Benford, is as follows: for base $b$, the digit $d \in \{1, 2, \ldots, b-1\}$ occurs with relative frequency

\[p_b(d) = \log_b \left(\frac{d+1}{d}\right) = \frac{1}{\log b}\log \left(\frac{d+1}{d}\right).\]

It initially seems bizarre that digits do not occur with equal frequency. But as neatly explained by Pietronero et al. (1998), it follows immediately if the relevant numbers are logarithmically uniform. Let $X$ be our random number. Then the first digit is $d$ if

\[db^k \leq X < (d+1)b^{k} \quad \Longrightarrow \quad \log_b d + k \leq \log_b X < \log_b(d+1) + k\]

for some integer $k$. If $X$ is logarithmically uniform, for instance sitting near the mean of a big random product, then $\log_b X$ is uniformly distributed, and lies in the interval $I_d := [\log_b d, \log_b (d+1)]$ with probability

\[(\log_b (d+1) + k) - (\log_b d + k) = \frac{1}{\log b}\log \left(\frac{d + 1}{d}\right) = p_b(d).\]

This provides a simple way to check for fraud on tax returns, for instance. Just compute relative frequencies of first digits in different bases and check they obey Newcomb-Benford! You might wonder why something totally deterministic, like the first digit of a power of $2$, also obeys Benford’s law. Here is a pie chart of initial decimal digits for the first $10,000$ binary powers, which follows the Newcomb-Benford law exactly:

Here is the Python code to generate it. You can check it for other numbers besides $2$ as well:

import matplotlib.pyplot as plt
import math

maxpower = 10000 # Number of powers to check
power = 2 # Change to check other powers

nums = '1', '2', '3', '4', '5', '6', '7', '8', '9',
benford = [(math.log(10, d+1) - math.log(10, d+1))
    for d in range(1, 10)]
firstdig = [0 for i in range(9)]
for i in range(maxpower):
	ind = int(str(power**i)[0]) - 1
	firstdig[ind] = firstdig[ind] + 1

fig, ax = plt.subplots()
ax.pie(firstdig, labels=nums, autopct='%1.1f%%', startangle=90)
    # Change 'firstdig' to 'benford' for probabilities

The mechanism for logarithmic uniformity here is slightly different, and discussed in depth in Serge Tabachnikov’s book on geometric billiards. In this case, $X = 2^n$, so the first digit is $d$ just in case

\[\log_{10}d + k \leq n\log_{10} 2 < \log_{10}(d + 1) + k.\]

Let $\text{frac}(x)$ denote the fractional part of $x$, and define $x_n := \text{frac}(n\log_{10} 2)$. Taking fractional parts gives

\[\log_{10}d \leq x_n < \log_{10}(d + 1).\]

It turns out that, since $x_1 = \log_{10} 2$ is irrational, $x_n$ jumps randomly around the unit interval, and forms an “equidistribution” which spends equal times in equal areas. For a proof, see Tabachnikov’s book. But although the fundamental cause is different, the outcome is still logarithmic uniformity, and the Newcomb-Benford law results.

The philosophy of subestimates

Now we’ve dealt with geometric means and logarithmic uniformity, we turn to the effectiveness of factorising a Fermi estimate. If we take logarithms, factors become summands, and we’ll reason about those since they are simpler. If $Z = X + Y$ is a sum of independent random variables, the variance is additive, so that

\[\text{var}(Z) = \text{var}(X) + \text{var}(Y).\]

Thus, splitting a sum into estimates of the summands and adding them should not change the variance of the guess. Of course, there is a fallacy in this reasoning: humans are not sampling from the underlying distribution! When we guess, we introduce our own random errors. For instance, my estimate for $Z$ will have some human noise $\varepsilon_Z$:

\[\hat{Z} = Z + \varepsilon_Z.\]

Similarly, my guesses for $X$ and $Y$ have some random errors $\varepsilon_X$ and $\varepsilon_Y$. There is no reason for the variances of $\varepsilon_X$ and $\varepsilon_Y$ to add up to the variance of $\varepsilon_Z$. The sum could be bigger, or it could be smaller. A judicious decomposition should reduce the combined variance:

\[\text{var}(\varepsilon_X) + \text{var}(\varepsilon_Y) < \text{var}(\varepsilon_Z).\]

If log-normality is the science of Fermi estimates, picking variance-reducing subestimates is the art. But there is a connection to our earlier discussion. I think the human error $\varepsilon_X$ will roughly mimic the empirical distribution of $Z$ we have seen in the world. If it is biased, so is $\varepsilon_X$; it we have only seen a few examples, the variance of $\varepsilon_X$ will probably be large, and decrease roughly as $1/k$ with $k$ examples. So the general strategy for variance reduction is to factorise into things we have seen before. We can even use these data points to generate subestimates by geometric averaging.

Variance reduction in practice

I’m not sure there is any particularly good rule of thumb for when to factor. A simple one, however, is as follows: try generating over- and underestimates for the factors and the product. In additive notation, go with the smaller of

\[(\Delta X)^2 + (\Delta Y)^2, \quad (\Delta Z)^2\]

where $\Delta$ refers to the difference of the (logarithm of the) over- and underestimate. Let’s illustrate by returning to the population of Chile. I can try factoring it into a number of regions multiplied by the average number of people per region. Taking logs (in base $10$) of the over- and underestimate of Chile’s population I gave above, I get

\[(\Delta Z)^2 = (\log_{10} 10^8 - \log_{10} 10^6)^2 = 4.\]

On the other hand, for regions I would make a lower guess of $5$ and an upper guess of $30$, with a difference in logs of $(\Delta X)^2 = 0.6$. For regional population, I would make a lower guess of $5\times 10^5$ and an upper guess of $10^7$, with $(\Delta Y)^2 = 1.7$. Thus,

\[(\Delta X)^2 + (\Delta Y)^2 = 2.3 < 4 = (\Delta Z)^2.\]

The guess from the factorisation (taking geometric means) is

\[\sqrt{5 \times 30 \times (5\times 10^5) \times 10^7} \approx 27 \text{ million}.\]

Logarithmically, this is indeed closer to the actual population than our earlier guess. For reference, the number of regions is $16$, while our mean is around $12$, and the average population is a bit over a million, while we’ve overestimated at $2.25$ million. But the two balance out and give a better overall estimate. This suggests a diversity of prediction mechanism is at play with subestimates, but I haven’t worked out the details.

All this suggests that, while there is considerable art, there is also some solid statistics underlying the effectiveness of Fermi estimates. They’re not so unreasonable after all!