Normal Distribution: From a Bell Shape to a Probability Model

A continuous density must be nonnegative, have total area one, and assign interval probability through area under the curve.

For a continuous variable, P(a < X < b) is the integral of f from a to b; a single height f(x) is density, not probability.

The kernel g(x)=exp(-x^2) has the desired bell shape, but its area is not yet one, so it is not yet a probability density.

If a nonnegative kernel has finite area A, dividing every height by A creates f(x)=g(x)/A with total area one.

Multiplying every height by the same positive constant preserves the curve's center, ordering, symmetry, and relative shape while changing its total area.

Both are symmetric, but x^2 grows in the tails and -x^2 becomes negative; neither can be normalized into a density on the whole real line.

Squaring gives left-right symmetry, a negative sign makes distance a penalty, and exponentiation keeps the result positive while making tails fade.

Points mu-d and mu+d have equal squared deviations, so a density based on (x-mu)^2 gives them equal height.

Mirrored inputs at the same distance from mu receive the same density; the square creates symmetry, while the exponential creates the bell's decay.

A larger squared standardized distance makes the exponent more negative, so exp(-z^2/2) remains positive but rapidly approaches zero.

At z=0,1,2,3 the unnormalized weights exp(-z^2/2) are approximately 1, 0.607, 0.135, and 0.011.

In exp(-c(x-mu)^2), a larger positive c penalizes distance faster and produces a narrower bell; a smaller c produces a wider bell.

The Normal chooses c=1/(2 sigma^2); increasing sigma decreases c, stretches the bell, and lowers its peak so total area stays one.

No. Many positive integrable functions normalize into symmetric bells, but only the exponential-quadratic family is Normal.

A Normal tail decays exponentially in squared distance, while a Cauchy tail decays polynomially; that difference changes moments and extreme-event behavior.

Differentiating exp(u(x)) returns u'(x)exp(u(x)), keeping the same exponential family without an extra logarithmic factor.

For a>1, a^(-c x^2)=exp(-(c ln a)x^2); changing the base only rescales the positive quadratic coefficient.

The product c ln(a) is the effective decay rate, so a new base can be offset by a new c or sigma without changing the exponential-quadratic family.

Mu locates the center, sigma controls spread, the square enforces symmetry, and the leading factor 1/(sigma sqrt(2 pi)) makes total area one.

Changing mu shifts the bell without changing shape; changing sigma adjusts width and peak height while preserving total area one.

Calling the unknown integral I and squaring it converts two identical one-dimensional integrals into a double integral over the plane.

With x^2+y^2=r^2 and dxdy=r dr dtheta, the angular integral contributes 2 pi and the radial integral contributes 1/2, so I^2=pi.

The product exp(-x^2)exp(-y^2)=exp(-(x^2+y^2)) is a height over the xy-plane; equal-height level sets are circles.

A thin ring of radius r and thickness dr has area approximately 2 pi r dr; integrating equal-height rings sums the surface volume.

Substituting x=z/sqrt(2) shows integral exp(-z^2/2) dz=sqrt(2 pi), so its reciprocal normalizes the standard Normal density.

Where differentiable, F'(x)=f(x); conversely F(x)=integral from negative infinity to x of f(t)dt, so differentiation localizes accumulated probability.

For any variable with finite positive standard deviation, Z=(X-mu)/sigma measures signed distance from the mean in standard-deviation units.

If X is Normal(mu,sigma^2), then subtracting mu and dividing by sigma produces Z Normal(0,1), enabling one reference curve and table.

For independent Normal X and Y, the sum is Normal; means add and variances add, preserving the family exactly.

A standardized sum of many independent identically distributed finite-variance effects approaches a standard Normal under the classical central limit theorem.

No. The classical theorem assumes independent identically distributed effects with finite nonzero variance; other CLTs use different technical conditions.

Some heavy-tailed distributions have undefined variance and fall outside the classical finite-variance CLT; their sums can have different limiting behavior.

The Cauchy density is symmetric and bell-shaped but has much heavier tails and no finite mean or variance, so ordinary z-score and finite-variance CLT reasoning fail.

Among continuous densities with a fixed finite mean and variance, the Normal has the largest differential entropy and adds the least extra distributional structure.

Use it when the mechanism, diagnostics, and purpose support symmetric light-tailed variation; check skew, outliers, bounds, mixtures, and dependence rather than assuming a bell.