What Is a Normal Distribution?
The word "normal" here is mathematical shorthand, not a claim that all natural data is bell-shaped. The distribution was formalized by Carl Friedrich Gauss in the early 19th century — which is why it's also called the Gaussian distribution — though Abraham de Moivre described the bell curve shape more than a century earlier in 1733 while working on approximations to the binomial distribution.
In practice, many physical measurements on large populations approximate the normal distribution because of a deeper theorem in probability: the Central Limit Theorem. When a variable is the sum of many small, independent random effects, its distribution converges toward normal regardless of the underlying distribution of those individual effects.
μ = mean (center of the distribution)
σ = standard deviation (spread)
σ² = variance
e ≈ 2.71828 (Euler's number)
π ≈ 3.14159
The PDF gives the relative likelihood of each value of x. You can't read a direct probability from it for a single point (since any single value on a continuous scale has probability zero). Instead, you find probabilities by computing the area under the curve between two values, which is why calculus — or a z-table — is involved in probability calculations.
The Bell Curve Shape
The "bell curve" nickname comes from the profile of the graph: a single rounded peak in the center that slopes downward on both sides, eventually flattening out as it approaches (but never quite reaches) the horizontal axis. The peak sits exactly at the mean.
Several features make the bell curve distinctive. The mean, median, and mode all fall at the same point: the top of the bell. The curve is perfectly symmetric about that center, so the left half is a mirror image of the right half. This symmetry means extreme high values are exactly as probable as extreme low values at the same distance from the mean.
The tails of the distribution extend infinitely in both directions. In theory, any value is possible — but the probabilities become vanishingly small beyond about 3 standard deviations from the mean. Less than 0.27% of a normally distributed population falls outside that range.
Key Properties of the Normal Distribution
- Symmetry: The distribution is perfectly symmetric about the mean. P(X < μ − k) = P(X > μ + k) for any value k.
- Mean = Median = Mode: All three measures of center coincide at μ. There is exactly one peak (unimodal).
- Asymptotic tails: The curve approaches but never touches the x-axis. The tails extend to −∞ and +∞.
- Total area = 1: The entire area under the curve integrates to exactly 1, consistent with a valid probability distribution.
The shape of the bell curve changes with σ. A small standard deviation produces a tall, narrow bell; a large standard deviation produces a short, wide one. The area under each curve remains 1 regardless of shape. The mean μ shifts the entire curve left or right along the horizontal axis without changing its shape.
The Empirical Rule (68-95-99.7 Rule)
The empirical rule is a shorthand for three probability statements that hold for every normal distribution, regardless of its specific mean and standard deviation. Understanding it makes rough probability calculations possible without any tables or calculators.
For any normally distributed variable: 68.27% of values fall within 1 standard deviation of the mean (μ ± σ); 95.45% fall within 2 standard deviations (μ ± 2σ); 99.73% fall within 3 standard deviations (μ ± 3σ). The rule is sometimes written as 68-95-99.7.
| Range | Interval | % of Data Included | % Outside Range |
|---|---|---|---|
| Within 1σ | μ − σ to μ + σ | 68.27% | 31.73% |
| Within 2σ | μ − 2σ to μ + 2σ | 95.45% | 4.55% |
| Within 3σ | μ − 3σ to μ + 3σ | 99.73% | 0.27% |
Applied to exam scores: if a test has μ = 70 and σ = 10, then 68.27% of students scored between 60 and 80. 95.45% scored between 50 and 90. A score below 40 or above 100 — three standard deviations away — would occur in fewer than 1 in 370 students. For a full treatment, see the empirical rule guide.
Standard Normal Distribution and Z-Scores
There are infinitely many normal distributions — one for every combination of μ and σ. Calculating probabilities for all of them separately would be impractical, so statisticians use a single reference distribution called the standard normal distribution, written Z ~ N(0, 1).
x = observed value
μ = population mean
σ = population standard deviation
z = standardized score (in units of σ)
The z-score answers this question: how many standard deviations is x away from the mean? A z-score of 1.5 means the value is 1.5 standard deviations above the mean. A z-score of −0.8 means it's 0.8 standard deviations below. Once you have z, you can look up the corresponding cumulative probability in a z-table or use the z-score calculator.
For worked examples using z-scores in detail, the z-score examples page covers eight fully solved problems across different domains. The complete reference is in the z-score guide.
10 Real-Life Normal Distribution Examples
Each of the following variables is either known to follow, or well-approximated by, a normal distribution in large populations. Where known population parameters exist (from census data, testing organizations, or published research), those are listed. Where they are estimates, the source is noted.
Standardized Test Scores
The College Board designed the original SAT so that each section would have a mean of 500 and a standard deviation of 100 — an explicit choice to create a normal distribution. This makes probability calculations straightforward: a score of 700 sits 2σ above the mean, meaning roughly 97.7% of test-takers score below it.
Adult Human Height
Height in adult males within a single population follows a close approximation to the normal distribution. CDC data from the National Health and Nutrition Examination Survey measured average male height in the US at 69.1 inches (5′9.1″) with σ ≈ 2.9 inches. Female heights are similarly distributed at μ = 63.6 in, σ = 2.8 in.
IQ Scores
IQ scores are designed to be normally distributed, not observed to be. The Wechsler Adult Intelligence Scale is calibrated so that μ = 100 and σ = 15 in the norming sample. This design means exactly 2.28% of the population scores above 130 (2σ above mean), which is the approximate Mensa threshold. The 68-95-99.7 rule applies directly.
Newborn Birth Weight
Full-term newborn birth weights approximate a normal distribution with mean around 7.5 pounds. Low birth weight (below 5.5 lb) corresponds to roughly z = −1.8, meaning about 3.6% of full-term infants — consistent with reported clinical rates. Pediatricians use this distribution to flag neonates at the tails for additional monitoring.
Adult Blood Pressure
Systolic blood pressure in a general adult population follows approximately a normal distribution. The American Heart Association sets the Stage 1 hypertension threshold at ≥130 mmHg — a z-score of about 0.71 on the population distribution, meaning roughly 24% of adults meet or exceed that threshold. Clinical screening programs use this to size testing needs.
Manufacturing — Bolt Diameters
Precision manufacturing produces dimensions that cluster around a target with small, normally distributed variation. A bolt specified at 10.00 mm diameter with σ = 0.05 mm means 99.73% of bolts fall within ±0.15 mm of the target. Six Sigma methodology directly uses this property: a "six sigma" process has its tolerance limits at ±6σ, leaving only 3.4 defects per million opportunities.
Daily High Temperatures
Daily maximum temperatures in a given month and location approximate a normal distribution when measured over many years. Chicago's July high temperature records show a mean around 83°F with standard deviation around 7°F. A heat wave day above 97°F corresponds to z = 2.0, with roughly 2.3% probability on any July day. Utility companies use this to plan peak power demand.
Employee Performance Ratings
Many organizations that use numerical performance ratings find the resulting distribution approximates normal when the scale is well-calibrated and raters are well-trained. Forced distribution systems (the old GE "vitality curve") explicitly target a normal shape. A rating of 4.4 or above would fall in roughly the top 2% on a distribution with μ = 3.0, σ = 0.7.
Annual Stock Market Returns
Annual returns on broad market indexes are often modeled as approximately normal for basic financial calculations. Historical S&P 500 data shows mean annual return around 10% with standard deviation around 18%. This gives a loss worse than −26% (z ≈ −2) in roughly 2.3% of years. Note: daily returns have fat tails, making the normal approximation less reliable at short time horizons.
Reaction Time in Psychology
Human reaction time on a simple stimulus-response task (press button when light flashes) follows an approximately normal distribution in controlled lab conditions. With μ ≈ 250 ms and σ ≈ 40 ms, a response faster than 170 ms corresponds to z = −2, occurring in roughly 2.3% of trials. Psychologists use this baseline to detect impairment from sleep deprivation, alcohol, or aging.
3 Fully Worked Examples with Calculations
The examples below follow a consistent four-step pattern: identify parameters, compute the z-score, find the probability from the standard normal distribution, and state the result in plain language. All z-table values come from the standard normal z-table.
Worked Example 1 — Student Exam Scores
Problem: A statistics professor grades 300 students' final exams. Scores are normally distributed with a mean of μ = 72 and a standard deviation of σ = 11. What percentage of students scored above 90? What percentage scored between 61 and 83?
Part A: P(X > 90)
Identify parameters: μ = 72, σ = 11, x = 90.
Compute z-score: z = (90 − 72) / 11 = 18 / 11 = 1.636
Round to z ≈ 1.64 for table lookup.
Look up probability: From the z-table, P(Z < 1.64) = 0.9495. Therefore: P(X > 90) = 1 − 0.9495 = 0.0505
✅ Answer: About 5.1% of students scored above 90. In a class of 300, that's approximately 15 students.
Part B: P(61 < X < 83)
Compute both z-scores:
z₁ = (61 − 72) / 11 = −11/11 = −1.00
z₂ = (83 − 72) / 11 = 11/11 = 1.00
Look up probabilities: P(Z < 1.00) = 0.8413 and P(Z < −1.00) = 0.1587.
Subtract: P(−1 < Z < 1) = 0.8413 − 0.1587 = 0.6827
This is the 68.27% from the empirical rule — 61 and 83 are exactly 1σ below and above the mean.
✅ Answer: About 68.3% of students scored between 61 and 83 — approximately 205 of the 300 students.
Worked Example 2 — Human Height
Problem: Adult male heights in the US follow N(μ = 69.1 in, σ = 2.9 in) based on CDC NHANES data. (a) What percentage of adult men are shorter than 5 feet 6 inches (66 inches)? (b) What height marks the 90th percentile?
Part A: P(X < 66)
Compute z-score: z = (66 − 69.1) / 2.9 = −3.1 / 2.9 = −1.069
Round to z ≈ −1.07.
Look up probability: P(Z < −1.07) = 0.1423.
✅ Answer: About 14.2% of adult men are shorter than 5′6″.
Part B: 90th percentile height
Find z for 90th percentile: From the z-table, P(Z < z) = 0.90 gives z ≈ 1.282.
Convert back to raw score: x = μ + z × σ = 69.1 + 1.282 × 2.9 = 69.1 + 3.72 = 72.82 inches (approximately 6 feet 0.8 inches).
✅ Answer: The 90th percentile height for adult men is approximately 72.8 inches (about 6 feet 1 inch). Only 10% of adult men are taller than this.
Worked Example 3 — Manufacturing Quality Control
Problem: A machine produces stainless steel pins with a target diameter of 5.00 mm. Measurements are normally distributed with μ = 5.00 mm and σ = 0.04 mm. Pins are rejected if their diameter falls outside the range 4.90 mm to 5.10 mm. What percentage of pins pass inspection?
Identify parameters: μ = 5.00, σ = 0.04, specification limits: 4.90 to 5.10 mm.
Compute z-scores for each limit:
z₁ = (4.90 − 5.00) / 0.04 = −0.10 / 0.04 = −2.50
z₂ = (5.10 − 5.00) / 0.04 = 0.10 / 0.04 = +2.50
Look up probabilities: P(Z < 2.50) = 0.9938 and P(Z < −2.50) = 0.0062.
Calculate the passing percentage:
P(4.90 < X < 5.10) = 0.9938 − 0.0062 = 0.9876 = 98.76%
✅ Answer: 98.76% of pins pass inspection. In a production run of 10,000 pins, approximately 124 would be rejected. To reduce defects further, the manufacturer could tighten σ (improve machine precision) or widen the tolerance (relax the spec). The relationship between σ and tolerance limits is the basis of process capability analysis.
Normal Distribution Calculator
Enter a mean, standard deviation, and a value (or range of values). The calculator returns the z-score and cumulative probability. Use it to verify the worked examples above or solve your own problems. For a standalone tool, see the normal distribution calculator and the interactive bell curve generator.
Normal Distribution Probability Calculator
How to Check if Your Data Is Normally Distributed
The assumption of normality is required by many statistical tests — t-tests, ANOVA, linear regression residuals. Before applying these tests, it's worth checking whether the data is plausibly normal. No real dataset is perfectly normal; the question is whether the deviation from normality matters given your sample size and analysis.
Plot a Histogram
The quickest first check. A roughly symmetric, unimodal (single-peaked) histogram consistent with the bell shape is encouraging. Strong skewness or multiple peaks rules out normality. For small samples (n < 30), histograms are unreliable — the shape is too noisy.
Q-Q Plot (Quantile-Quantile Plot)
A Q-Q plot graphs observed quantiles against theoretical normal quantiles. Points hugging the diagonal line indicate normality. Systematic curves or S-shapes suggest skewness or heavy tails. This is more sensitive than a histogram for detecting departures from normality.
Shapiro-Wilk Test
The most powerful formal test for normality for samples under 2,000. The null hypothesis is that the data was drawn from a normal distribution. A p-value above 0.05 means you fail to reject normality. For large n, the test becomes sensitive to tiny, practically meaningless deviations — which is a limitation. Full details on all formal tests are at the normality tests guide.
Descriptive Statistics — Skewness and Kurtosis
A normal distribution has skewness = 0 and kurtosis = 3 (or excess kurtosis = 0). Values close to these are consistent with normality. Skewness beyond ±1 or excess kurtosis beyond ±2 warrants concern. These statistics are reported by every major statistical software package alongside descriptive output. See the descriptive statistics guide for formulas.
With n > 1,000, formal normality tests will almost always reject the null hypothesis because tiny, practically irrelevant deviations become statistically detectable. For large samples, rely on Q-Q plots and the robustness of your chosen statistical procedure rather than formal tests alone.
Comparison Tables
Normal Distribution vs Standard Normal Distribution
| Feature | Normal Distribution | Standard Normal Distribution |
|---|---|---|
| Mean (μ) | Any real number | 0 |
| Standard deviation (σ) | Any positive number | 1 |
| Notation | X ~ N(μ, σ²) | Z ~ N(0, 1) |
| Probability calculation | Requires conversion to z first | Direct z-table lookup |
| Total area under curve | 1 | 1 |
| Conversion | z = (x − μ) / σ | Already standardized |
| Tables available | No — must standardize first | Yes — z-table |
Normal Distribution vs Other Common Distributions
| Distribution | Shape | Mean vs Median | When to Use |
|---|---|---|---|
| Normal | Symmetric bell curve | Mean = Median = Mode | Measurements near a natural target with random error |
| Right-skewed | Long right tail | Mean > Median > Mode | Income, house prices, reaction times with outliers |
| Left-skewed | Long left tail | Mode > Median > Mean | Exam scores near a ceiling, age at death |
| Uniform | Flat rectangle | Mean = Median | Random number generation, dice rolls |
| Binomial | Discrete, bell-like for large n | Mode ≈ Mean | Count of successes in n trials |
| T-distribution | Like normal but fatter tails | Mean = Median = 0 | Small samples with unknown σ — see t vs normal |
Where the Normal Distribution Is Used in Practice
Clinical Medicine
Reference ranges for lab tests (cholesterol, hemoglobin, glucose) are set using the normal distribution. "Normal" reference intervals typically cover μ ± 2σ, meaning roughly 95% of healthy patients fall inside.
Quality Control
Statistical Process Control charts monitor manufacturing processes by tracking whether measured dimensions stay within ±3σ of target. A point outside those limits signals a process shift. This is the foundation of Six Sigma methodology.
Finance & Risk
Value at Risk (VaR) models — required by banking regulators — frequently assume normally distributed returns to estimate the worst expected loss at a given confidence level. The assumption works reasonably for annual returns but fails for daily returns where fat tails are observed.
Scientific Research
Residuals from regression models should follow a normal distribution for inference to be valid. Researchers check normality as a standard part of model diagnostics before reporting coefficients and confidence intervals. See: hypothesis testing guide.
Standardized Testing
IQ tests, SAT, GRE, and professional licensing exams are calibrated using item response theory to produce score distributions that approximate the normal. Score reports express performance in terms of percentiles derived directly from normal distribution calculations.
Machine Learning
Gaussian Naive Bayes classifiers assume features follow a normal distribution within each class. Gaussian mixture models extend this to datasets with multiple clusters. Feature normalization (z-score standardization) transforms any approximately normal feature to Z ~ N(0,1) for use in algorithms sensitive to scale.
Common Mistakes When Working with Normal Distributions
| Mistake | What People Say | What's Actually True |
|---|---|---|
| Assuming everything is normal | All large datasets follow a normal distribution | Many do not. Incomes, city populations, and website traffic are log-normal or power-law distributed. Check first. |
| Confusing N(μ,σ²) notation | N(100, 15) means mean=100, SD=15 | The second parameter is the variance (σ²), not σ. N(100, 225) means μ=100, σ=15. Notation varies by textbook — always check. |
| Reading P(X = x) from a z-table | P(X = 72) = the z-table value at z=0 | For continuous distributions, P(X = any exact value) = 0. Z-tables give cumulative probabilities: P(Z < z). |
| Forgetting to check normality | I have 500 data points so normality is fine | Large n ensures CLT applies to sample means, not that raw data is normal. Skewed raw data remains skewed at any n. |
| Misreading the tail | P(Z > 1.5) is the z-table value at z=1.5 | Most z-tables give P(Z < z). For a right-tail probability: P(Z > 1.5) = 1 − P(Z < 1.5) = 1 − 0.9332 = 0.0668. |
Key Terms Glossary
| Term | Symbol | Definition |
|---|---|---|
| Normal Distribution | N(μ, σ²) | Symmetric, bell-shaped continuous probability distribution defined by mean and variance |
| Gaussian Distribution | — | Alternate name for the normal distribution, honoring Carl Friedrich Gauss |
| Mean | μ | The expected value; center of the normal distribution |
| Standard Deviation | σ | Square root of variance; measures spread. Larger σ = wider bell |
| Variance | σ² | Average squared deviation from the mean; the second parameter of N(μ, σ²) |
| Z-Score | z | Number of standard deviations a value is from the mean: z = (x − μ)/σ |
| Standard Normal | Z ~ N(0,1) | Normal distribution with μ=0 and σ=1; used for all z-table lookups |
| f(x) | Probability density function; gives relative likelihood at each value of x | |
| CDF | F(x) or Φ(z) | Cumulative distribution function; gives P(X ≤ x) |
| Empirical Rule | 68-95-99.7 | ~68% of data within 1σ; ~95% within 2σ; ~99.7% within 3σ of the mean |
| Percentile | — | The value below which a given percentage of the distribution falls |
| Skewness | γ₁ | Measure of asymmetry. Normal distribution has skewness = 0 |
| Kurtosis | γ₂ | Measure of tail heaviness. Normal distribution has kurtosis = 3 (excess kurtosis = 0) |
| Central Limit Theorem | CLT | States that sample means approach a normal distribution as n grows, regardless of the population's distribution |
Frequently Asked Questions
A normal distribution is a way data spreads out around an average. Most values land near the middle, with progressively fewer values as you move further away in either direction. The result, when graphed, looks like a symmetrical bell shape. Human heights are a classic example: most adults are close to average height, with fewer people who are very tall or very short.
The graph of the normal distribution's probability density function looks like the cross-section of a bell. It starts low on the left, rises to a single rounded peak in the center (at the mean), then descends symmetrically on the right back to near zero. The shape has no sharp corners or flat sections — just the smooth, rounded profile of a bell.
Variables that approximate a normal distribution include human height, birth weight, blood pressure, IQ (by design), standardized test scores, errors in physical measurements, daily temperature averages, and dimensions of manufactured parts. Annual stock returns are sometimes modeled as normal for rough calculations, though the actual distribution has fatter tails. The key requirement is that the variable results from many independent additive factors — which is why the Central Limit Theorem makes normality so common in practice. See the full list in the 10 examples section.
They refer to the same distribution. "Normal distribution" is the standard statistical term used in textbooks and software. "Gaussian distribution" is the name used in physics, engineering, and machine learning, honoring Carl Friedrich Gauss who described it in his work on astronomical measurement errors in the early 1800s.
Many hypothesis tests rely on the normal distribution for their theoretical justification. The one-sample z-test assumes the test statistic follows a standard normal distribution under the null hypothesis. The t-test uses the t-distribution (which converges to normal as sample size grows). Even chi-square and F distributions can be approximated using normal distribution results for large degrees of freedom. The hypothesis testing guide covers each test in detail.
The normal distribution assumption breaks down for data that is heavily skewed (incomes, website traffic, city populations), bounded below zero (counts, durations), binary (yes/no outcomes), multimodal (two or more distinct peaks), or rare events (insurance claims, equipment failures). For skewed positive data, the log-normal distribution is often a better fit. For count data, the Poisson or negative binomial distributions are more appropriate. For binary outcomes, the binomial or logistic model applies. Always plot your data before choosing a distributional assumption.
The total area under a normal distribution curve equals exactly 1 (or 100%), because it represents all possible outcomes. This is a requirement for any valid probability distribution. The area between two values represents the probability that a randomly selected observation falls in that range — which is how z-table probabilities are defined.
A z-score and its corresponding percentile are directly linked through the cumulative distribution function of the standard normal. A z-score of 0 corresponds to the 50th percentile (the median). Z = 1.00 corresponds to approximately the 84th percentile (P(Z < 1) ≈ 0.8413). Z = −1.645 corresponds approximately to the 5th percentile. The z-score to percentile conversion tool on the visual tools page handles this automatically.
Sources and References
All probability values in this guide are calculated from the standard normal CDF, cross-checked against the NIST Engineering Statistics Handbook. Height and weight parameters are from published CDC anthropometric data.
- NIST Engineering Statistics Handbook — Normal Distribution. National Institute of Standards and Technology. itl.nist.gov
- OpenStax Introductory Statistics — Ch. 6: The Normal Distribution. Rice University. openstax.org
- Fryar, C.D. et al. (2016) — Anthropometric Reference Data for Children and Adults: United States, 2011–2014. CDC Vital and Health Statistics, Series 3, No. 39. cdc.gov
- Penn State STAT 500 — Applied Statistics. Lesson 4: Probability Distributions. Penn State Eberly College of Science. online.stat.psu.edu
- Khan Academy — Normal Distributions Review. khanacademy.org
- Gauss, C.F. (1809) — Theoria Motus Corporum Coelestium in Sectionibus Conicis Solem Ambientium. Hamburg: Perthes et Besser. First mathematical formalization of the normal distribution in the context of measurement error.