What is a Gaussian distribution?

In this article

1. Before you start
2. Why interviewers ask this
3. Gaussian distribution from first principles
4. Worked example
5. Trade-offs and edge cases
6. How to answer it in an interview
7. Common mistakes
8. Follow-up questions
9. Sources

Before you start

Why interviewers ask this

A candidate who cannot explain why the bell curve appears will tend to apply it to quantities where it fails, and one follow-up shows it. The Gaussian is a natural first distribution to ask about because it tests three skills at once: reading a formula, reasoning about averages, and doing quick probability arithmetic without a calculator. Beyond "define the normal distribution", the questions in our collection for this topic include "why does it show up", "what is the probability that one standard normal exceeds three times another" and "build two Gaussians with a given correlation".

A weak answer recites the formula. A strong answer says what the two parameters mean, why averages are Gaussian, and where the shape stops being the right model.

Gaussian distribution from first principles

In plain words

Gaussian distribution

A Gaussian distribution is a probability distribution over real numbers that is fully described by its mean and its standard deviation and has a symmetric bell-shaped density. It is also called the normal distribution.

Standard deviation

The standard deviation is a number that measures how far values typically sit from their mean, in the same unit as the values. Its square is the variance.

z-score

A z-score is the number of standard deviations a value lies above or below the mean, so it turns any Gaussian value into one on a common scale.

Central limit theorem

The central limit theorem is a result that says the average of many independent draws from one distribution is approximately Gaussian, whatever the shape of the original distribution.

Think of weighing one coin a thousand times on a kitchen scale. Each reading is the true weight plus many tiny errors. Some push up and some push down. Most readings land near the middle and few land far away. That pile of readings is bell-shaped.

The analogy breaks when the pushes are not many, small and added together. A waiting time is not a true value plus small errors. At a constant rate, short waits are the most common and long ones trail off. So waiting times are skewed, not bell-shaped. Being unable to go negative is not the reason: the coin readings cannot go negative either.

In one precise sentence, a Gaussian has a mean that sets where the curve is centred and a standard deviation that sets how wide it is. Everything else in this page, from the averaging result to the z-score, follows from those two numbers and the symmetry of the curve.

Step 1: Two numbers describe the whole curve

The Gaussian density is fixed by two parameters, the mean $\mu$ and the standard deviation $\sigma$. The NIST normal distribution page writes the density as

and calls $\mu$ the location parameter and $\sigma$ the scale parameter. The input is the pair $(\mu ,\sigma )$. The output is a curve that peaks at $\mu$ and is symmetric around it.

Changing $\mu$ slides the curve left or right without changing its shape. Changing $\sigma$ stretches or squeezes it. SciPy uses the same split: its loc argument is the mean and its scale argument is the standard deviation, and the standard normal is the case $\mu =0$, $\sigma =1$ (SciPy 1.18.0 reference).

Step 2: What the central limit theorem actually says

The central limit theorem says the distribution of the sample average, not of the raw data, approaches a Gaussian as the sample size grows. OpenStax, central limit theorem for sample means states that for samples of size $n$ from a population with mean $\mu$ and standard deviation $\sigma$, the sample means tend to be normally distributed with mean $\mu$ and standard deviation $\sigma /\sqrt{n}$, the standard error of the mean.

The raw data keeps its own shape however many rows you collect. A column of waiting times stays skewed at a million rows. What becomes Gaussian is the average of repeated samples of it.

The input is a source distribution with a finite mean and variance. The output is a spread that shrinks as $\frac{1}{\sqrt{n}}$. In a simulation of 200,000 averages of flat (uniform) draws, the share inside plus or minus 1.96 theoretical standard deviations went from 100.0% at $n=1$ to 95.0% at $n=30$¹.

Step 3: A z-score turns any Gaussian into the standard one

A z-score $z=(x-\mu )/\sigma$ converts a value from any Gaussian to the standard normal, so one table of probabilities serves every Gaussian. OpenStax, the standard normal distribution defines the z-score as how many standard deviations a value is above or below the mean.

The same source gives the empirical rule: about 68% of values lie within one standard deviation of the mean, about 95% within two and about 99.7% within three. Computed exactly, the shares are 68.27%, 95.45% and 99.73%, and the central 95% sits at $\pm 1.96$ standard deviations¹.

The input is a raw value plus the mean and standard deviation. The output is a distance in standard deviations, which you can read as a tail probability.

Step 4: Sums and symmetry, or why two Gaussians split a plane evenly

Independent Gaussians stay Gaussian when added, and their variances add, so the standard deviation of a sum grows like a square root. The sum of two independent standard normals has standard deviation $\sqrt{2}\approx 1.414$; a simulation of one million sums gave 1.4144¹.

Two independent standard normals $(X,Y)$ form a point whose density depends only on the distance from the origin, so the cloud is a circle with no preferred direction. Any wedge holds the share of the probability that its angle covers of a full turn. The event $Y>3X$ is a half-plane through the origin, so its probability is exactly one half.

To build two standard normals with correlation $\rho$, take two independent ones $Z_1,Z_2$ and set $A=Z_1$, $B=\rho Z_1+\sqrt{1-{\rho }^2}{Z}_2$. With $\rho =0.6$, a simulation gave a correlation of 0.5996 and a variance of 0.9979 for $B$¹.

1 · Two parameters2 · Averages are Gaussian3 · The z-score4 · SymmetryReplay

Step-by-step description

1. The mean sets the centre and the standard deviation sets the width.
2. The sample average, not the raw data, approaches a Gaussian.
3. Standardising puts every Gaussian on one scale.
4. A circular cloud splits evenly, so Y > 3X has probability one half.

• The Gaussian density is set by the mean and the standard deviation.
• The sample average has standard deviation sigma over the square root of n.
• Within one, two and three standard deviations lie 68.27%, 95.45% and 99.73% of values.
• For independent standard normals the probability that Y exceeds 3X is exactly one half.

Worked example

The spread is the standard deviation. The input is 200,000 trials per row of averages of uniform(0, 1) draws, each with theory standard deviation $\sqrt{1/(12n)}$. Excess kurtosis is a number that measures tail weight relative to a Gaussian, where zero means Gaussian tails. It rises from −1.197 toward zero as $n$ grows, which is the shape converging. With one draw averaged, all of the uniform values fall inside the interval, 100.0%; with thirty draws averaged, 95.0% of the averages land in the 95% interval¹.

Text description of the diagram

• With thirty draws averaged, 95.0% of the averages land in the 95% interval.
• With one draw averaged, all of the uniform values fall inside the interval, 100.0%.
• Each row used 200,000 trials.

The conditional case works the same way. Given $X>0$ the point is in the right half of the plane, and $Y>3X$ there is a wedge from the line $y=3x$ (at 71.57 degrees) up to the vertical axis, an angle of 18.43 degrees. That is $\frac{18.43}{360}=0.0512$ of the whole plane, and dividing by the half-plane's share of 0.5 gives 0.1024. A simulation of two million pairs gave 0.4996 and 0.1025¹.

A z-score example, with illustrative numbers: a test with mean 100 and standard deviation 15 and a score of 130 gives $z=2.0$, which sits at the 97.72nd percentile if scores are Gaussian¹.

Trade-offs and edge cases

The Gaussian is the right model for sums and averages, and the wrong one for many other quantities. The time until an event that happens at a constant rate is skewed and fits an exponential; the number of events in a fixed time interval fits a Poisson; a positive quantity whose logarithm is Gaussian fits a log-normal, which is what a product of many positive factors tends to give. The NIST handbook gives each its own page and formula: the exponential distribution is used to model data with a constant failure rate, the Poisson distribution the number of events in a given time interval, and the lognormal distribution is defined by a normal logarithm.

Text description of the diagram

• The time until an event that happens at a constant rate is skewed and fits an exponential.
• The number of events in a fixed time interval fits a Poisson.
• A positive quantity whose logarithm is Gaussian fits a log-normal.

The central limit theorem has conditions and no universal sample size. It needs independent draws and a finite variance. The folk rule "thirty is enough" has no basis in the theorem itself: the OpenStax section states the result for large sample sizes and gives no minimum $n$. Our flat source is symmetric and converged fast; a heavily skewed source would need more draws.

In machine learning the Gaussian appears in two more places. Weight initialisation in deep learning often draws starting weights from Gaussians, and a common noise assumption behind linear regression in classical machine learning is Gaussian. It also appears as a loss. If the noise around a prediction is Gaussian with fixed variance, the negative log-likelihood is proportional to the squared error, so minimising squared error, which gradient descent does step by step, is maximum likelihood under that assumption. The matching result for classification is the cross-entropy loss, and the entropy, cross-entropy and Kullback-Leibler (KL) divergence page builds the idea from the ground up.

How to answer it in an interview

The spoken structure is: define it with two parameters, give the reason it appears, then say where it stops applying.

Common mistakes

• "As you collect more data, the data becomes normal" (wrong). The raw data keeps its shape; the distribution of the sample average becomes normal (right).
• "Thirty samples is always enough" (wrong). The theorem gives no minimum, and skewed sources need more draws (right).
• "Standard deviation and variance are the same number" (wrong). Variance is the square of the standard deviation and has squared units (right).
• "P(Y > 3X) must be small because 3X is big" (wrong). The event is a half-plane through the origin, so it is exactly one half for independent standard normals (right).
• "Add standard deviations when adding independent variables" (wrong). Variances add, so the standard deviations combine as a square root of a sum of squares (right).

Follow-up questions

What is the central limit theorem?

The sample average of many independent draws is approximately Gaussian with mean $\mu$ and standard deviation $\sigma /\sqrt{n}$, whatever the source shape, as long as the variance is finite. It explains why the Gaussian appears in measurement error, polling and A/B tests.

How do I build two standard normals with correlation rho from independent ones?

Take independent $Z_1$ and $Z_2$, set $A=Z_1$ and $B=\rho Z_1+\sqrt{1-{\rho }^2}{Z}_2$. Then $B$ has variance one and the correlation of $A$ and $B$ is $\rho$. This is the one-step case of the Cholesky construction used for larger covariance matrices from linear algebra.

What is the probability that Y is greater than 3X for independent standard normals?

It is exactly one half, because the density is rotation-symmetric and the event is a half-plane through the origin. Given $X>0$ the answer drops to about 0.1024.

Which distribution other than the normal should I know, and where does it apply?

The Poisson distribution models the number of events in a given time interval, such as requests per second. It is set by one parameter, its mean, and its standard deviation is the square root of that mean, so its variance equals its mean, which is a quick check on whether count data fits (NIST Poisson page).

What happens to the spread of an average if the sample size quadruples?

The standard error is halved, because it scales as $\frac{1}{\sqrt{n}}$. Cutting it by a factor of ten needs a hundred times more data, which is why precision gets expensive.

Sources

1. Introductory Statistics, section 6.1: The Standard Normal Distribution — OpenStax (Illowsky and Dean), 19 Sep 2013
2. Introductory Statistics 2e, section 7.1: The Central Limit Theorem for Sample Means — OpenStax, 13 Dec 2023
3. NIST/SEMATECH e-Handbook of Statistical Methods, 1.3.6.6.1 Normal Distribution (page undated; read 2026-10-24) — NIST, 24 Oct 2026
4. NIST/SEMATECH e-Handbook of Statistical Methods, 1.3.6.6.7 Exponential Distribution (page undated; read 2026-10-24) — NIST, 24 Oct 2026
5. NIST/SEMATECH e-Handbook of Statistical Methods, 1.3.6.6.9 Lognormal Distribution (page undated; read 2026-10-24) — NIST, 24 Oct 2026
6. NIST/SEMATECH e-Handbook of Statistical Methods, 1.3.6.6.19 Poisson Distribution (page undated; read 2026-10-24) — NIST, 24 Oct 2026
7. SciPy v1.18.0 reference: scipy.stats.norm (read 2026-10-24) — SciPy developers, 24 Oct 2026