Complement
The complement of an event is the event that it does not happen, consisting of every outcome in the sample space that is not in the original event, and its probability is exactly 1 minus the probability of the event. Written P(not A) = 1 - P(A), this complement rule follows directly from the fact that an event and its complement together cover the entire sample space without overlap, so their probabilities must sum to 1. It is written A^c, A' or the event A with a bar over it, and it is one of the shortest and most useful calculation shortcuts in all of probability.
The complement rule is valuable because the opposite of a hard question is often an easy one, especially for "at least one" problems. The probability of rolling at least one six in five rolls of a fair die would require summing many overlapping cases directly, but its complement, no sixes at all, has probability (5/6)^5, about 0.402, so the answer is 1 - 0.402 = 0.598. The method scales effortlessly: at least one defective item in a batch, at least one server failure across a cluster, at least one false positive across many tests. Whenever the target phrase is "at least one" or "not all," the strategy is to compute the probability of "none" and subtract.
The birthday problem is the canonical showcase of the complement rule. Asking for the probability that at least two people in a group of 23 share a birthday (ignoring leap days and assuming equally likely birthdays) looks intractable, yet the complement requires only multiplying 365/365 by 364/365 by 363/365 and so on down to 343/365, giving about 0.4927 for all-different and therefore about 0.507 for at least one match. The surprise, that 23 people suffice for better-than-even odds, comes from the number of pairs, 253, growing quadratically with group size. The same calculation underlies hash-collision analysis in computing, where the birthday bound determines how many items can be hashed before an accidental collision becomes likely.
Complements interact with the other set operations through De Morgan's laws, named for the 19th-century mathematician Augustus De Morgan, and misapplying them is the most common error. The complement of (A or B) is (not A) and (not B), and the complement of (A and B) is (not A) or (not B). Hence the opposite of "at least one of the sensors fails" is "all of the sensors work," not "all fail," and the opposite of "all are defective" is "at least one is not defective." In reliability, a parallel system with n independent components, each failing with probability q, fails only if all fail, so its reliability is 1 - q^n; a series system works only if every component works. Writing out the complement in words and checking it against these laws avoids the most frequent slip.
The complement rule has practical limits, and several pitfalls are worth remembering. It requires a correctly specified sample space: the complement is relative to the whole space, so if the space is wrong, the answer is wrong. For continuous quantities, an event of probability zero has a complement of probability one without the complement covering literally every outcome. The rule says nothing by itself about independence; computing the chance of "none" as a simple product such as (5/6)^5 requires the trials to be independent, and if they are not (as with draws without replacement), the product must use conditional probabilities. In statistics, the p-value, confidence level and power are linked by complements: a 95 percent confidence level is 1 minus a 5 percent significance level, and power is 1 minus the Type II error rate.
Read the complement as the shortcut that lets a hard question be answered by its easy opposite: whenever an event is awkward to count directly, compute the probability that it fails to happen and subtract from 1, taking care that the opposite of "at least one" is "none" and that the opposite is always measured within the same, correctly defined sample space.