10
Part Zero · Foundations

Probability and statistics

On 4 July 2012, physicists at CERN announced that they had found the Higgs boson. They had seen a small bump in a graph of collision data, and bumps appear by pure chance all the time. They waited to announce until the chance of a fluke as large as theirs was below about one in three and a half million. Established

Why that number? Quantum mechanics only ever predicts probabilities, and every measurement comes with noise, so the language of chance is not optional in this book. This chapter gives you enough of it to judge a claim of discovery yourself.

Average and spread

Roll a fair die many times. The expectation, the long-run average, is ⟨x⟩=∑x P(x)=3.5\langle x\rangle = \sum x\,P(x) = 3.5, a value the die never shows. The variance measures spread: the average squared distance from the mean, σ2=⟨x2⟩−⟨x⟩2\sigma^2 = \langle x^2\rangle - \langle x\rangle^2. Its square root, the standard deviation σ\sigma, is in the same units as xx.

Quantum mechanics uses exactly these: the expectation value of a measurement and its uncertainty Δx\Delta x, which is a standard deviation.

Why bell curves are everywhere

A Galton board drops balls through rows of pegs. At each peg a ball bounces left or right at random. through 12 rows.

Each ball’s final position is a sum of many small random steps, and the heap they build is a bell curve. That is the central limit theorem: add up enough independent random contributions, whatever their individual shape, and the total is approximately Gaussian. Measurement errors, thermal jiggling and noise in detectors are sums of this kind, which is why the Gaussian appears in every lab.

Counting rare events

A detector clicks when a particle passes, at a steady average rate but at random moments. The number of clicks in a fixed window follows the Poisson distribution, P(n)=μne−μ/n!P(n) = \mu^n e^{-\mu}/n!, where μ\mu is the average count. Set it to 3.0.

Its standard deviation is μ\sqrt\mu. Count 100 events and the natural wobble is about 10; count 10,000 and it is about 100. Relative uncertainty falls as 1/μ1/\sqrt{\mu}, which is why experiments run for years. For large μ\mu the Poisson shape becomes a Gaussian, as the central limit theorem says it must.

The base-rate trap

A disease affects 1.0 % of people. A test catches 90 % of real cases and wrongly flags 9.0 % of healthy people. You test positive. How likely is it that you are ill?

Most people, including many doctors in surveys, say about 90%. The plate shows a thousand people. The true answer is 9%. With a rare disease, the many false alarms among the healthy swamp the few true cases.

This is Bayes’ theorem: P(ill ∣ +)=P(+ ∣ ill) P(ill)/P(+)P(\text{ill}\,|\,+) = P(+\,|\,\text{ill})\,P(\text{ill})/P(+). It is also how physicists should think about a surprising result: how rare was the effect before you looked?

Sigmas and discoveries

Particle physicists measure surprise in standard deviations. If a bump is nσn\sigma above the expected background, the chance that background alone would fluctuate that high is the area of the Gaussian tail beyond nn. Set 3.0 σ: the chance is 1 in 741.

Three sigma is called “evidence”; five, about one in 3.5 million, is the bar for “discovery”. The bar is high because experiments look at many places at once, so a one-in-a-thousand fluke is expected somewhere: the look-elsewhere effect. In December 2015 both big LHC experiments reported an excess near 750 GeV in pairs of photons; it was a few sigma at its peak and vanished with more data in 2016. Established

Fitting a line

Given data with Gaussian errors, which line fits best? Ask which line makes the observed data most probable: maximize the likelihood. For Gaussian errors that is the same as minimizing the sum of squared misses, χ2=∑(yi−f(xi))2/σi2\chi^2 = \sum (y_i - f(x_i))^2/\sigma_i^2. This is least squares. and watch the fitted line wobble around the true one.

DeeperThe likelihood is ∏i12πσiexp⁡ ⁣(−(yi−f(xi))22σi2)\prod_i \frac{1}{\sqrt{2\pi}\sigma_i}\exp\!\left(-\frac{(y_i - f(x_i))^2}{2\sigma_i^2}\right). Taking the logarithm turns the product into a sum, and maximizing it is minimizing χ2\chi^2. For a straight line f=a+bxf = a + bx, setting ∂χ2/∂a=∂χ2/∂b=0\partial\chi^2/\partial a = \partial\chi^2/\partial b = 0 gives two linear equations for aa and bb.

What you now have

Expectation and spread, the central limit theorem, Poisson counting, Bayes’ theorem and its trap, significance in sigmas, and least squares. Every experimental claim in this book can be read with these.

· · ·

Three questions to carry with you.

Why is the Gaussian so common?tap
The central limit theorem: sums of many independent random contributions are approximately Gaussian, whatever the shape of each one.
Why can a positive result from a good test still mean you are probably healthy?tap
If the condition is rare, false positives from the large healthy group outnumber the true positives.
Why do physicists demand 5σ rather than 3σ?tap
Many places are searched at once, so 3σ flukes are expected somewhere; 5σ (about 1 in 3.5 million) protects against the look-elsewhere effect and hidden errors.

And one to work out. With prevalence 1%, sensitivity 90% and a 9% false-positive rate, what is P(ill ∣ +)P(\text{ill}\,|\,+), as a fraction?

P =

Sources: ATLAS Collaboration, Phys. Lett. B 716, 1 (2012); CMS Collaboration, Phys. Lett. B 716, 30 (2012). G. Gigerenzer, Calculated Risks (2002). R. J. Barlow, Statistics. Checked by code: 5σ one-sided p = 2.87×10⁻⁷; P(ill|+) = 0.092.

Next · Chapter 11
Newtonian mechanics and conservation laws

Why are orbits ellipses and not some other shape?

Turn the page