Why randomness creates streaks and clusters

Six heads looks unusual. Its probability is not unique.

Compare two exact sequences: HHHHHH and HTHTHT. For a fair coin with independent flips, each has probability (½)⁶, or 1/64 = 1.5625%. The alternating sequence is just as specific as the run of heads.

The comparison changes when one side is a whole category. “All heads” contains one sequence. “Any mixture of heads and tails” contains 62. A larger category can be more likely even though every individual sequence has the same chance.

After six heads, the next independent fair flip is still 50% heads and 50% tails. Tails does not become due. The model contains no mechanism that compensates for what happened earlier.

That does not prove a real coin or generator is fair. It tells us what follows if the process is fair and independent. Evaluating an unknown process requires evidence about how it works.

A longer window gives a streak more places to appear

“Will the next five flips all be heads?” is different from “Will five heads appear in a row anywhere during the next twenty flips?” The first asks about one fixed block. The second allows a run to begin in several places.

At 20 flips, the probability is 24.987%, compared with 3.125% for the next five flips being all heads. At 100 flips, it rises to about 81.011%. A longer observation window changes how surprising a run is without changing the coin.

There is a catch: the possible starting positions overlap. A sequence of six heads contains two five-head windows. Adding their probabilities as though they were separate outcomes counts some sequences more than once.

We counted sequences containing no five-head run, then subtracted their share from one. An independent exhaustive check found that exactly 262,008 of the 1,048,576 twenty-flip sequences contain a qualifying run. The reproducible calculation is linked in the methodology.

Randomness leaves room for both clusters and gaps

Imagine placing 64 dots on a rectangle. Under a uniform random model, each dot gets an independently chosen horizontal and vertical coordinate. Every equal-area region has the same expected number of dots, but a particular sample need not divide them evenly.

If we divide the placement area into 16 equal regions, each region has an expected count of 64 ÷ 16 = 4 dots. In the displayed random sample, the actual counts range from one to seven. In the constructed grid, every region contains exactly four.

Neither picture, by itself, establishes whether an unknown generator is fair. A cluster can occur under the random model; a tidy result can come from a rule that deliberately prevents clustering. The placement process is what makes the comparison meaningful.

A deck changes when a card is removed

Independent coin flips leave the next-flip probability unchanged. Drawing a card without replacing it changes the available population. A standard deck begins with four Aces among 52 cards, so the first-card chance of an Ace is 4/52, or about 7.69%.

Remove an Ace and the next-card Ace probability falls to 3/51, about 5.88%. Remove any non-Ace and it rises to 4/51, about 7.84%. The remaining deck has changed, so the probability changes with it.

There is no contradiction between “tails is not due” and “the next Ace probability depends on which cards are gone.” Those statements describe different processes.

The population matters as much as the rule. Spider Solitaire 2 Suits uses 104 cards, with eight Aces across its two suits. Before any cards are dealt, that pack’s Ace share is also 1/13, but removing one Ace leaves 7/103, not 3/51. The one-deck calculation cannot simply be carried over after a removal; a probability for the stock during play must also account for cards already known elsewhere on the table.

Information matters too. In FreeCell, every card is dealt face up: the shuffled layout sets the puzzle, but there is no face-down stock to draw from. Uncertainty about how to solve that layout is a different problem from uncertainty about an unseen card. These examples describe information and card counts, not game win rates.