Probability — page 2

A walk that samples a distribution

When a distribution can be evaluated but not drawn from, a wandering point can be arranged to visit each state as often as its weight says. The rule needs no normalising constant, compares two weights and steps or stays.

Points too even to be random

Independent random points clump, and the clumping is what makes the error fall only as the square root. Points chosen to be evenly spread rather than independently beat that rate, and the price is that nothing about them is random at all.

The chain that stops

Give a chain a state it cannot leave and there is no long run to find — every walk ends. What is worth computing instead is how long it lasts and where it finishes, and both are exact answers to a linear system rather than limits of anything.

The chain that runs the same backwards

Put weights on the edges of a graph, step to a neighbour in proportion to them, and the long-run share of a state is its own weight over the total — read straight off the picture, with nothing to solve. The condition that makes that work is strictly stronger than being stationary.

The time spent and the share held

Stationary shares are a limit of distributions — where the walk probably is after many steps. Here the question is about a single walk: the fraction of its time spent in each state is that state's share, and the expected wait between visits is exactly the reciprocal.

Where the shares have nowhere to go

On finitely many states, a chain that can reach everywhere and is not forced into a rhythm settles down. Give it infinitely many and both conditions can hold while the walk leaves and never returns — or returns with certainty and takes an unbounded average time about it.

How long until it forgets

The essays before this one settle where a chain ends up and how much time it spends there, and none of them asks how long the settling takes. That question has an exact answer, it is a single number, and it is the only thing any practical use of a chain depends on.

The moment a giant appears

Raise the chance of an edge slowly and a random graph does nothing for a long time, then in a narrow window acquires a component holding a definite fraction of everything. The fraction is the root of an equation, and the equation says why the transition is where it is.

Two thresholds, not one

A random graph acquires a piece holding most of its points at average degree one, and is still not connected. Connectivity waits until the average degree reaches the logarithm of the size, and what holds it up is the very last isolated point.

Finding a threshold with two moments

Every monotone property of a random graph has a threshold, and locating one is nearly always the same two calculations — count what the property needs, and check the count does not concentrate on rare cases. The triangle is where the method is cleanest.

The window where the giant is born

Below the threshold the largest piece is a few dozen points, above it a definite fraction of everything. At the threshold it is neither, and the size it does take — the two-thirds power — is an exponent with no elementary derivation that a measurement finds immediately.

Sharp, or merely a threshold

Every monotone property of a random graph has a threshold. Some of them turn on over a range that shrinks relative to the threshold as the graph grows, and some do not — and which kind a property is turns out to be decided by whether it is about a local structure or about the whole graph.

When the numbers are shown

The secretary rule wins a third of the time and cannot do better, because it is told only who is ahead. Show the actual values and say where they came from, and the same problem is won three times in five — by a standard that falls as the end approaches.

Giving up on the best

The secretary rule treats landing the second-best exactly as badly as landing the worst, which is a strange thing to want. Ask instead for the smallest average rank and the answer is about the fourth-best candidate — whatever the size of the field, and whether it is ten or ten million.

Half of what an oracle takes

Compare an online rule not against the best it could have done but against a rule that has seen every value in advance. One fixed threshold secures half of what the oracle collects, whatever the distributions are — and there is an example on which half is all there is.

The thresholds that nest

Allow a second acceptance in the secretary problem and the chance of holding the best rises from about 37 per cent to about 59. The best rule is still a threshold — but one threshold for each number of choices still in hand, the earlier ones starting sooner, and each additional choice buying less than the one before.

Add the odds from the end

Watch a sequence of independent events and try to stop exactly on the last one that happens. Add up the odds of the events from the end backwards until the total reaches one, and stop at the first success from there. That rule is the best possible for any probabilities whatever, and the secretary problem is the special case in which the chances are one over the position.

A sum stopped early still says something

Inclusion–exclusion corrects an overcount, then the correction's overcount, and so on to the end. Stop after any number of terms and the result is not merely an approximation: after an odd number it is too high and after an even number too low, always. So two or three terms bracket an answer whose full sum is out of reach — as long as the events being counted are rare.

How many get their own hat

The chance that nobody gets their own hat settles on 1/e. The chance that exactly one person does settles on 1/e too, exactly two on 1/(2e), exactly three on 1/(6e) — the Poisson distribution with mean 1. The reason is a set of averages that come out exactly 1 at every size, and the counts reach the limit so fast that eight hats are within six ten-thousandths of it.

The bound is the answer to a search

Chebyshev's inequality is not a clever estimate that happens to be sharp. It is the exact answer to a maximisation over all distributions with a stated mean and variance, and the polynomial that proves nothing beats it is the certificate a search of that kind always produces.

No single input can move it far

Independence was never the hypothesis doing the work. A quantity built from many separately drawn inputs concentrates whenever changing one of them moves it only a little — and that covers quantities which are not sums of anything and have no formula at all.

The one that hardly ever comes up

Make the kinds unequally likely and the tidy decomposition into stages fails, because a stage's rate now depends on which kinds turned up rather than on how many. What replaces it is an alternating sum over every subset — and the rarest kind turns out to be nearly the whole answer.

Two patterns, one chance, different waits

HTH and HTT are equally likely in any given window of three tosses. Waiting for HTH takes ten tosses on average and waiting for HTT takes eight, and the difference is not about probability at all — it is about what a failed attempt leaves behind.

Any unevenness brings the match sooner

Real birthdays are not spread evenly across the year, and every such departure pushes the famous twenty-three down rather than up. The proof is one move on two days at a time, and what it leaves behind is a single number — the one ecologists use to count species.