FINITE MATHEMATICS • PROBABILITY AND STATISTICS

Normal Approximation

Using the continuous normal curve to efficiently estimate discrete probabilities when exact computation becomes impractical.

Historical Context & Motivation

The story of the normal approximation is deeply intertwined with the development of probability theory itself. In the eighteenth century, mathematicians working on games of chance frequently needed to compute binomial probabilities — sums of the form C(n, k) pk(1 − p)n−k — for large values of n. Without electronic calculators, evaluating factorials like 100! or summing dozens of individual terms was an extraordinarily tedious task. The central question became: is there a smooth, continuous function that closely mirrors the shape of the binomial distribution when n is large, allowing us to replace exact summation with a simple integral?

1733
De Moivre's Discovery
Abraham de Moivre showed that the binomial distribution for a fair coin (p = ½) could be approximated by an exponential curve — essentially the bell curve — when the number of trials is large. His work appeared in a privately circulated pamphlet and later in The Doctrine of Chances.
1809
Gauss and the Error Curve
Carl Friedrich Gauss derived the normal distribution independently while analyzing astronomical measurement errors. His formulation of the least-squares method relied on the assumption that errors follow a bell-shaped distribution, giving the curve its alternate name — the Gaussian distribution.
1812
Laplace's Generalization
Pierre-Simon Laplace extended de Moivre's result to arbitrary values of p (not just ½) in his Théorie analytique des probabilités. This generalization is now known as the De Moivre–Laplace theorem, the foundational result behind the normal approximation to the binomial.
1922
Continuity Correction Formalized
Statisticians in the early twentieth century, including Karl Pearson and his school, systematized the continuity correction (adding or subtracting 0.5) to improve the accuracy of the normal approximation when bridging discrete and continuous distributions.

The fundamental question these mathematicians confronted remains the same one we address today: when the number of trials n is large, how can we efficiently estimate discrete probabilities without summing an impractical number of terms? The normal approximation provides the answer by replacing a discrete probability mass function with the elegant, well-tabulated normal density curve.

Core Principles & Definitions

The normal approximation rests on several interconnected ideas from probability theory. Understanding these principles clarifies both when the approximation is valid and why it works so remarkably well. At its core, the technique leverages the fact that many discrete distributions become increasingly bell-shaped as their parameters grow, a phenomenon guaranteed by the Central Limit Theorem.

1

De Moivre–Laplace Theorem

As the number of trials n increases, the standardized binomial distribution B(n, p) converges in distribution to the standard normal N(0, 1). This theorem is the theoretical backbone of the normal approximation to the binomial.
2

Continuity Correction

Because the normal distribution is continuous while the binomial is discrete, we adjust the boundary of our interval by ±0.5 to account for the width of each probability bar. This correction substantially improves accuracy for moderate n.
3

Rule-of-Thumb Conditions

The approximation is generally reliable when both np ≥ 5 (or 10) and n(1 − p) ≥ 5 (or 10). These conditions ensure the distribution is sufficiently symmetric and bell-shaped for the normal curve to be a good fit.
4

Standardization (Z-Score)

To use standard normal tables, we convert the binomial variable X to a Z-score via Z = (X − μ) / σ, where μ = np and σ = √(np(1 − p)). This maps the problem into the standard normal framework.
KEY TAKEAWAY
Think of the normal approximation as replacing a bar chart with a smooth curve drawn over it. When you have a few bars (small n), the smooth curve is a poor fit — it misses the staircase shape. But when the bar chart has hundreds of narrow bars, the smooth curve hugs it so tightly that the area under the curve is nearly identical to the total height of the bars. The continuity correction is like extending each bar to fill its full unit-width slot, ensuring the curve captures the entire bar rather than just its center.

Visual Explanation

The following diagram illustrates the essence of the normal approximation. A binomial distribution with n = 30 and p = 0.4 is displayed as a bar chart, and the corresponding normal curve N(μ = 12, σ² = 7.2) is superimposed. Notice how the continuous curve closely follows the tops of the discrete bars, demonstrating why the area under the curve serves as an excellent estimate of the sum of bar heights.

The violet bars represent the exact binomial probabilities P(X = k) for B(30, 0.4). The cyan curve is the normal density N(12, 7.2). The close alignment between the two demonstrates why the area under the normal curve provides an excellent approximation to the sum of discrete probabilities.

Observe that the peak of the normal curve occurs at μ = np = 12, precisely where the tallest binomial bar is located. The spread of the curve is governed by σ = √(np(1 − p)) ≈ 2.68. Because both np = 12 and n(1 − p) = 18 comfortably exceed the threshold of 5 (or even 10), the fit between the discrete bars and the continuous curve is very tight. This visual correspondence is the geometric justification for substituting a normal probability calculation in place of an exact binomial summation.

Mathematical Framework

The mathematical machinery behind the normal approximation centers on three computations: identifying the parameters of the matching normal distribution, standardizing the variable, and applying the continuity correction. We formalize each step below.

BINOMIAL MEAN AND STANDARD DEVIATION
μ = np σ = √(np(1 − p))
where n is the number of independent trials and p is the probability of success on each trial. These are the mean and standard deviation of the binomial random variable X ~ B(n, p).
STANDARDIZATION (Z-SCORE)
Z = (X − μ) / σ = (X − np) / √(np(1 − p))
The Z-score measures how many standard deviations X lies from the mean. Under the normal approximation, Z is treated as a standard normal variable N(0, 1), allowing us to look up probabilities in the Z-table.
CONTINUITY CORRECTION
P(a ≤ X ≤ b) ≈ P((a − 0.5 − μ)/σ ≤ Z ≤ (b + 0.5 − μ)/σ)
When approximating P(a ≤ X ≤ b) for a discrete variable X, we expand the interval to [a − 0.5, b + 0.5] before standardizing. For P(X ≥ a), use a − 0.5; for P(X ≤ b), use b + 0.5; for P(X = k), use [k − 0.5, k + 0.5].
⚠️ Validity Conditions
Before applying the normal approximation to a binomial distribution B(n, p), verify that both np ≥ 5 and n(1 − p) ≥ 5. Some texts use a stricter threshold of 10. If either condition fails, the binomial distribution may be too skewed for the symmetric normal curve to approximate it well, and exact methods or other approximations (e.g., the Poisson) should be preferred.

The theoretical justification for these formulas comes from the De Moivre–Laplace theorem, which states that as n → ∞, the standardized binomial variable (X − np) / √(np(1 − p)) converges in distribution to a standard normal random variable. In practice, the convergence is rapid enough that for moderate n the approximation error is small — typically within a few percentage points of the exact probability, and often much better when the continuity correction is applied.

Continuity Correction in Detail

The continuity correction is arguably the most error-prone step for students learning the normal approximation. Because a discrete random variable assigns positive probability to individual integers while a continuous variable assigns zero probability to any single point, we need a systematic rule for translating discrete probability statements into continuous ones. The key insight is that each integer k in the binomial distribution occupies a conceptual 'bin' stretching from k − 0.5 to k + 0.5 on the real number line.

The highlighted amber bar at X = 12 spans the interval [11.5, 12.5] on the continuous axis. The continuity correction captures the full area under the normal curve (cyan) corresponding to this bar, ensuring the approximation accounts for the discrete bar's width.
Continuity correction translations for common probability statements
Discrete StatementCorrected Continuous IntervalRationale
P(X = k)P(k − 0.5 ≤ Y ≤ k + 0.5)Capture the full width of the single bar at k.
P(X ≤ k)P(Y ≤ k + 0.5)Include the bar at k by extending to its right edge.
P(X ≥ k)P(Y ≥ k − 0.5)Include the bar at k by extending to its left edge.
P(X < k)P(Y ≤ k − 0.5)Exclude the bar at k by stopping at its left edge.
P(X > k)P(Y ≥ k + 0.5)Exclude the bar at k by starting at its right edge.
P(a ≤ X ≤ b)P(a − 0.5 ≤ Y ≤ b + 0.5)Include both endpoint bars by expanding to their outer edges.

A useful mnemonic is: if the inequality includes the integer k (i.e., ≤ or ≥), then adjust by 0.5 in the direction that expands the interval. If the inequality excludes k (i.e., < or >), adjust by 0.5 in the direction that shrinks the interval. This ensures you capture exactly the bars you intend to count.

Worked Example

A quality-control engineer inspects a production run of 200 circuit boards. Historical data shows that 8% of boards contain a defect. Find the probability that at most 20 boards are defective using the normal approximation with continuity correction.

Normal Approximation to B(200, 0.08)
1
Step 1 — Identify the DistributionLet X = number of defective boards. Each board is inspected independently with P(defect) = 0.08, so X ~ B(n = 200, p = 0.08). We want P(X ≤ 20).
2
Step 2 — Check Validity ConditionsCompute np = 200 × 0.08 = 16 and n(1 − p) = 200 × 0.92 = 184. Both values exceed 5 (and even the stricter threshold of 10), so the normal approximation is appropriate.
np = 16 ≥ 10 ✓ n(1 − p) = 184 ≥ 10 ✓
3
Step 3 — Compute Mean and Standard Deviationμ = np = 200 × 0.08 = 16. σ = √(np(1 − p)) = √(200 × 0.08 × 0.92) = √14.72 ≈ 3.837.
μ = 16, σ ≈ 3.837
4
Step 4 — Apply Continuity CorrectionWe seek P(X ≤ 20). Since the inequality includes 20, we expand the interval to include the full bar at 20: P(X ≤ 20) ≈ P(Y ≤ 20.5), where Y ~ N(16, 3.837²).
Corrected boundary: 20.5
5
Step 5 — StandardizeZ = (20.5 − 16) / 3.837 = 4.5 / 3.837 ≈ 1.173.
Z ≈ 1.17
6
Step 6 — Look Up ProbabilityUsing the standard normal table, Φ(1.17) ≈ 0.8790. Therefore, P(X ≤ 20) ≈ 0.8790, or about 87.9%. For comparison, the exact binomial probability (computed via software) is approximately 0.8854, demonstrating the close agreement.
P(X ≤ 20) ≈ 0.879

Strengths, Limitations & When to Use Alternatives

Like any approximation technique, the normal approximation has a domain of reliable applicability and situations where it breaks down. Understanding these boundaries is essential for making sound modeling decisions in applied statistics.

Strengths and limitations of the normal approximation
StrengthsLimitations
Converts combinatorial sums into a single Z-table lookup, drastically reducing computation.Accuracy degrades when p is close to 0 or 1, since the binomial becomes highly skewed.
Works well for a wide range of n and p values, provided np ≥ 5 and n(1 − p) ≥ 5.For small n, the discrete staircase shape is too coarse for a smooth curve to match.
Generalizes beyond the binomial — applicable to sums of i.i.d. random variables via the CLT.Tail probabilities (far from the mean) are less accurate; the normal curve may overestimate or underestimate.
The continuity correction is straightforward to apply and significantly improves accuracy.Modern computing makes exact binomial CDF calculation fast, reducing the practical need for approximation.
🔄 WHEN TO USE ALTERNATIVES
When n is large but p is very small (e.g., p = 0.01), the binomial distribution is strongly right-skewed and the normal curve is a poor fit. In such cases, the Poisson approximation (with λ = np) is often more appropriate. Conversely, when n is small, no continuous approximation is suitable — use exact binomial calculations or software. Think of it as choosing the right lens: the normal lens works for the wide middle ground, the Poisson lens works for rare events, and the exact calculation is the high-resolution microscope for small samples.

Connection to the Central Limit Theorem

The normal approximation to the binomial is, in fact, a special case of one of the most profound results in all of probability theory: the Central Limit Theorem (CLT). The CLT asserts that the standardized sum (or mean) of n independent, identically distributed random variables — regardless of the underlying distribution — converges in distribution to N(0, 1) as n → ∞. Since a binomial random variable X ~ B(n, p) can be written as the sum of n independent Bernoulli trials X = X₁ + X₂ + ⋯ + Xn, the De Moivre–Laplace theorem is the CLT applied to Bernoulli random variables.

Normal approximation as a special case of the CLT
FeatureNormal Approximation (This Lesson)Central Limit Theorem (General)
Underlying distributionBinomial (sum of Bernoulli trials)Any distribution with finite variance
Key parametersμ = np, σ = √(np(1 − p))μ = nμ_X, σ = σ_X√n
Continuity correctionRequired (discrete → continuous)Only if summing discrete r.v.s
Typical applicationsQuality control, polling, binomial testsConfidence intervals, hypothesis testing, estimation
Rate of convergenceDepends on skewness of Bernoulli(p); fastest at p = 0.5Governed by the Berry–Esseen theorem: O(1/√n)

Understanding this connection positions you to extend the normal approximation idea far beyond the binomial setting. In later coursework on inferential statistics, you will use the CLT to construct confidence intervals and perform hypothesis tests for sample means drawn from virtually any population distribution — the same normal bell curve that approximates the binomial here will serve as the workhorse of modern statistical inference. The Berry–Esseen theorem provides a quantitative bound on the approximation error, guaranteeing that the maximum discrepancy between the true CDF and the normal CDF decreases at a rate proportional to 1/√n.

Practice Problems

PROBLEM 1CONCEPTUAL
Explain why the continuity correction adds 0.5 when computing P(X ≤ 15) but subtracts 0.5 when computing P(X > 15). What happens if you forget the correction — does your answer overestimate or underestimate the true probability?
PROBLEM 2BASIC CALCULATION
A fair coin is flipped 100 times. Use the normal approximation with continuity correction to estimate P(X ≥ 55), where X is the number of heads.
PROBLEM 3INTERMEDIATE
A survey reports that 35% of college students prefer online exams. In a random sample of 150 students, use the normal approximation to find P(45 ≤ X ≤ 60), where X is the number who prefer online exams.
PROBLEM 4APPLIED
A pharmaceutical company tests a new drug on 500 patients. The drug is expected to produce a positive response in 72% of patients. The company considers the trial successful if at least 350 patients respond positively. Using the normal approximation, estimate the probability of a successful trial.
PROBLEM 5CRITICAL THINKING
Consider X ~ B(n, p) where n = 1000 and p = 0.002. Explain why the normal approximation is inappropriate here despite n being very large. Propose a better approximation and compute P(X ≥ 5) using it. Then show what the normal approximation would give, and comment on the discrepancy.

Summary

The normal approximation allows us to estimate probabilities for a binomial distribution B(n, p) by replacing it with a continuous normal distribution N(μ, σ²) where μ = np and σ = √(np(1 − p)). This technique is justified by the De Moivre–Laplace theorem, a special case of the Central Limit Theorem. Before applying the approximation, always verify the validity conditions: both np ≥ 5 and n(1 − p) ≥ 5.

The key procedural steps are: compute μ and σ, apply the continuity correction (±0.5) to translate the discrete probability statement into a continuous interval, then standardize to a Z-score and look up the result in a standard normal table. When np is too small, consider the Poisson approximation instead. Mastery of this technique prepares you for the broader applications of the CLT in statistical inference, including confidence intervals and hypothesis testing.

Varsity Tutors • Finite Mathematics • Normal Approximation