Law of truly large numbers

The law of truly large numbers (a statistical adage), attributed to Persi Diaconis and Frederick Mosteller, states that with a large enough number of independent samples, any highly implausible (i.e. unlikely in any single sample, but with constant probability strictly greater than 0 in any sample) result is likely to be observed. Because we never find it notable when likely events occur, we highlight unlikely events and notice them more. The law is often used to falsify different pseudo-scientific claims; as such, it is sometimes criticized by fringe scientists.

The law can be rephrased as "large numbers also deceive". More concretely, skeptic Penn Jillette has said, "Million-to-one odds happen eight times a day in New York" (population about 8,000,000).

Examples
For a simplified example of the law, assume that a given event happens with a probability for its occurrence of 0.1%, within a single trial. Then, the probability that this so-called unlikely event does not happen (improbability) in a single trial is 99.9% (0.999).

For a sample of only 1000 independent trials, however, the probability that the event does not happen in any of them, even once (improbability), is only 0.9991000 ≈0.3677, or 36.77%. Then, the probability that the event does happen, at least once, in 1000 trials is (1 −0.9991000 ≈0.6323, or) 63.23%. This means that this "unlikely event" has a probability of 63.23% of happening if 1000 independent trials are conducted. If the number of trials were increased to 10,000, the probability of it happening at least once in 10,000 trials rises to (1 −0.99910000 ≈0.99995, or) 99.995%. In other words, a highly unlikely event, given enough independent trials with some fixed number of draws per trial, is even more likely to occur.

For an event X that occurs with very low probability of 0.0000001% (in any single sample, see also almost never), considering 1,000,000,000 as a "truly large" number of independent samples gives the probability of occurrence of X equal to 1 &minus; 0.9999999991000000000 ≈ 0.63 = 63% and a number of independent samples equal to the size of the human population (in 2021) gives probability of event X: 1 &minus; 0.9999999997900000000 ≈ 0.9996 = 99.96%.

These calculations can be formalized in mathematical language as: "the probability of an unlikely event X happening in N independent trials can become arbitrarily near to 1, no matter how small the probability of the event X in one single trial is, provided that N is truly large."

For example, where the probability of unlikely event X is not a small constant but decreased in function of N, see graph.

In high availability systems even very unlikely events have to be taken into consideration, in series systems even when the probability of failure for single element is very low after connecting them in large numbers probability of whole system failure raises (to make system failures less probable redundancy can be used - in such parallel systems even highly unreliable redundant parts connected in large numbers raise the probability of not breaking to required high level).

In criticism of pseudoscience
The law comes up in criticism of pseudoscience and is sometimes called the Jeane Dixon effect (see also Postdiction). It holds that the more predictions a psychic makes, the better the odds that one of them will "hit". Thus, if one comes true, the psychic expects us to forget the vast majority that did not happen (confirmation bias). Humans can be susceptible to this fallacy.

Another similar manifestation of the law can be found in gambling, where gamblers tend to remember their wins and forget their losses, even if the latter far outnumbers the former (though depending on a particular person, the opposite may also be true when they think they need more analysis of their losses to achieve fine tuning of their playing system ). Mikal Aasved links it with "selective memory bias", allowing gamblers to mentally distance themselves from the consequences of their gambling by holding an inflated view of their real winnings (or losses in the opposite case – "selective memory bias in either direction").