From a sum to a probability
A count of successes is discrete, but adding many independent trials produces fluctuations whose scale is the square root of the number of trials. The central limit theorem (CLT) describes the shape of those fluctuations. To use it, we must first identify the mean and standard deviation, translate the event correctly, and remember that an approximation is not an equality.
Let be independent Bernoulli variables with the same fixed success probability , where . Put and . Then , and .
Theorem 23.1.1
(Normal limit for binomial counts)
If is fixed, then for each fixed real ,
where is the standard normal distribution function.
This is the iid finite-variance CLT applied to the Bernoulli summands. Each has mean and variance ; independence removes the cross-covariances, so , and subtracting and dividing by produces a variable with mean zero and variance one. The theorem asserts convergence at fixed thresholds; it does not provide a numerical error bound for a particular . The fixed- assumption matters: if changes with , this stated iid theorem alone does not justify the conclusion. The degenerate cases are deterministic and cannot be standardized by this formula.
Integer endpoints and continuity correction
For integers , the exact binomial probability is
with terms outside taken as zero.
Definition 23.1.2
(Continuity correction)
A normal approximation with continuity correction treats the mass at integer as if it occupied the interval from to . Thus, writing ,
Likewise uses the lower boundary , while uses .
The half-unit adjustment is a useful approximation, not a theorem that makes the normal area exact or always more accurate. It is applied before standardization, since half a count becomes on the standardized scale.
Common mistake
(Placing the half-unit correction)
The half unit is added or subtracted on the count scale, before dividing by ; adding to a standardized value instead would move the boundary by counts. Its direction follows the inequality: and use , while and use . For , the events and therefore use and , although both are written with the number .
For a fixed , the CLT also supports boundaries that move toward a fixed standardized limit.
Proposition 23.1.3
(Thresholds converging to a fixed limit)
Let be fixed and . If real numbers satisfy , then . In particular, corrected and uncorrected normal approximations have the same limit.
Proof
(Sandwich between fixed thresholds)
Fix . For all large we have , so monotonicity gives
By the theorem, the outer terms tend to and , so the lower and upper limits of the middle term lie between these values. Letting and using continuity of gives . Adding half a count shifts a standardized boundary by , which is the special case.
Proof X-ray
Where continuity of the limit is used
The argument uses three facts, each once. Monotonicity of a distribution function in its threshold converts the moving threshold into two fixed ones; this needs only that is eventually within of , not any rate. The theorem is then applied at the fixed points , which is all it provides. The closing step is continuity of at : it makes the gap vanish as .
For a limit law with an atom at that step fails. In the Poisson regime of the counterexample below, the limiting distribution function jumps at every integer, so moving a threshold across an integer by a vanishing amount can change the limit. Nothing in the proof controls the speed of convergence, so it gives no finite-sample error estimate and no relative accuracy in extreme tails.
Worked example 23.1.4
(An interval of successes)
Suppose . The mean is and standard deviation . For , the corrected endpoints are and , so
The exact finite binomial sum is approximately . For the different event , use : the normal approximation is , while the exact sum is about . The agreement in these examples does not guarantee similar accuracy for all parameters or tails.
Worked example 23.1.5
(A one-sided tail with and without correction)
Let , so the mean is and the standard deviation is . For the corrected boundary is , giving
Ignoring the correction uses and gives . The exact binomial tail is approximately . The half unit matters at the second significant figure here, because it moves the standardized boundary by in a region where the normal density is still appreciable.
Worked example 23.1.6
(A single value needs the correction)
For , the uncorrected normal approximation assigns probability zero to every single value, since it subtracts at the same standardized point. The corrected approximation spreads the atom at over :
while the exact value is approximately . For point events the correction is therefore not a cosmetic refinement: without it the approximation is useless. It still carries no guarantee; it is accurate here partly because is the mean.
When the model or scale changes
Counterexample mode
A changing success probability
Take independent Bernoulli variables in row , each with . Then . The binomial law instead approaches a Poisson law of mean one. Its nonvanishing atom at zero cannot be captured by treating the count as approximately continuous merely because is large. Here stays at one, and the fixed-distribution CLT above is inapplicable.
Common mistake
(Copies of one trial are not independent trials)
If all trials are copies of one Bernoulli variable , the count is and has variance , not . A large sample size by itself does not supply independent information.
Similarly, a CLT approximation is not a probability bound: a bound holds under its stated assumptions, whereas an approximation requires a separate error analysis to quantify its accuracy.
Exercises
Quick check
(Let . Write a continuity-corrected normal approximation to and evaluate it to three decimals.)
Solution · Answer
The mean is and standard deviation . The approximation is .
Quick check
(For , give corrected normal expressions for and . Explain why their sum is one.)
Solution · Answer
The answers are and . Since is integer-valued, the events and partition the sample space, and both approximations use the boundary .
Quick check
(If with fixed , find the limit of , interpreting the real threshold literally.)
Solution · Answer
Standardization makes the event , so the CLT gives . No rounding is needed for this limiting statement.