Continuous Random Variables and Densities

A continuous random variable takes values in a continuum (an interval of $R$ ) and is described not by point probabilities — which are all zero — but by a probability density function (PDF) $f (x)$ whose integral over a set gives the probability of landing in that set. Key result: $P (a \leq X \leq b) = \int_{a}^{b} f (x) d x$ , with $f (x) \geq 0$ and $\int_{- \infty}^{\infty} f (x) d x = 1$ . Applies wherever measurements vary smoothly — physics, statistics, engineering, finance. Formalized by Laplace, Gauss, and rigorously by Kolmogorov (1933).

A random variable $X$ is continuous if its cumulative distribution function $F (x) = P (X \leq x)$ is continuous (and, for the density to exist, absolutely continuous), so that no single value carries positive probability. Instead, probability is spread out and measured by area under a density curve $f (x) = F^{'} (x)$ . This matters because most physical quantities — heights, lifetimes, voltages, arrival times — are modeled as continua, and densities give the cleanest language for expectation, variance, and transformations. The surprising fact: although $P (X = c) = 0$ for every individual point $c$ , the variable still certainly takes some value — probability zero does not mean impossible. The density $f (x)$ is itself not a probability and can exceed $1$ ; only its integral is constrained.

The Core Insight

For continuous variables, probability is area, not height. The density $f (x)$ answers "how concentrated is probability near $x$ ?" — formally $f (x) d x \approx P (x \leq X \leq x + d x)$ . You never read a probability off the curve directly; you always integrate. This single shift — from summing point masses (discrete case) to integrating a density (continuous case) — is the entire conceptual jump.

Formal Statement

Continuous Random Variable and Density

A random variable $X$ on a probability space $(Ω, F, P)$ is (absolutely) continuous if there exists a nonnegative integrable function $f : R \to [0, \infty)$ , the probability density function (PDF), such that for every $a \leq b$ :

P (a \leq X \leq b) = \int_{a}^{b} f (x) d x .

⟳This is the defining property; it asserts that the measure

P_{X}

induced by

X

is given by integrating

f

against Lebesgue measure.

The density must satisfy the two normalization axioms:

f (x) \geq 0 \forall x, \int_{- \infty}^{\infty} f (x) d x = 1.

⟳Nonnegativity follows since probabilities are nonnegative; the total integral equals

P (- \infty < X < \infty) = P (Ω) = 1

The cumulative distribution function (CDF) is

F (x) = P (X \leq x) = \int_{- \infty}^{x} f (t) d t,

⟳Set

a \to - \infty

b = x

in [1].

and where $F$ is differentiable the density is recovered by the Fundamental Theorem of Calculus:

f (x) = \frac{d}{d x} F (x) = F^{'} (x) .

⟳Differentiate [3]; FTC applies wherever

f

is continuous at

x

Symbols:

X

= random variable;

ω \in Ω

= sample point;

f

= density (units: inverse of

X

's units);

F

= CDF (dimensionless, in

[0, 1]

);

x, t

= real values.

Point Probabilities Vanish

If $X$ is continuous with density $f$ , then for every $c \in R$ ,

P (X = c) = 0,

and consequently

P (a \leq X \leq b) = P (a < X < b) = P (a \leq X < b) = P (a < X \leq b) .

⟳

P (X = c) = \int_{c}^{c} f = 0

by [1]; since endpoints contribute zero area, open/closed intervals share the same probability.

Expectation and Variance

For a continuous $X$ with density $f$ , and a measurable $g$ for which the integral converges absolutely (Law of the Unconscious Statistician):

E [g (X)] = \int_{- \infty}^{\infty} g (x) f (x) d x .

⟳Push the abstract expectation

\int_{Ω} g (X) d P

forward to the line via the distribution

P_{X} = f d x

; equivalently approximate

g

by simple functions.

In particular the mean $μ$ and variance $σ^{2}$ are

μ = E [X] = \int_{- \infty}^{\infty} x f (x) d x, σ^{2} = \int_{- \infty}^{\infty} (x - μ)^{2} f (x) d x = E [X^{2}] - μ^{2} .

⟳Take

g (x) = x

and

g (x) = (x - μ)^{2}

in [7]; expand the square and use linearity for the shortcut form.

Why It Holds / Derivation

From Axioms to the Density Picture

We show that [1]–[4] are mutually consistent and that the change-of-variable formula for transformations follows.

Step 1 — CDF is nondecreasing and bounded. Since

f \geq 0

, the integral

F (x) = \int_{- \infty}^{x} f

is nondecreasing. As

x \to - \infty

F \to 0

; as

x \to + \infty

F \to \int_{- \infty}^{\infty} f = 1

by [2]. Thus

F

is a legitimate CDF. Step 2 — Interval probabilities. By additivity of the integral,

P (a \leq X \leq b) = \int_{a}^{b} f = \int_{- \infty}^{b} f - \int_{- \infty}^{a} f = F (b) - F (a),

recovering [1] in terms of

F

. Step 3 — Recovering $f$ . Where

f

is continuous, FTC gives

F^{'} (x) = f (x)

, which is [4]. Heuristically,

P (x \leq X \leq x + Δ x) = F (x + Δ x) - F (x) \approx f (x) Δ x,

the "density times width" interpretation. Step 4 — Transformation rule. Let

Y = u (X)

with

u

strictly monotonic and differentiable, inverse

x = u^{- 1} (y)

. For increasing

u

F_{Y} (y) = P (u (X) \leq y) = P (X \leq u^{- 1} (y)) = F_{X} (u^{- 1} (y))

. Differentiating with the chain rule, and handling decreasing

u

with an absolute value, yields the change-of-variables density

f_{Y} (y) = f_{X} (u^{- 1} (y)) \frac{d}{d y} u^{- 1} (y) .

⟳Differentiate the CDF composition

F_{Y} = F_{X} \circ u^{- 1}

; the absolute value compensates for orientation reversal when

u

decreases.

■

Andrey Kolmogorov | 1903–1987

Kolmogorov's 1933 monograph Grundbegriffe der Wahrscheinlichkeitsrechnung placed probability on measure-theoretic foundations. In his framework a continuous random variable is a measurable function whose distribution is absolutely continuous with respect to Lebesgue measure; the density $f$ is precisely the Radon–Nikodym derivative $d P_{X} / d λ$ . This dissolved century-old ambiguities about "infinitely small probabilities" and explained why some continuous-CDF distributions (the Cantor distribution) have no density at all.

Building Intuition

Imagine spreading exactly one kilogram of fine sand along a number line. The density $f (x)$ is the mass per unit length at position $x$ — measured in kg/meter, not kg. To find how much sand lies between $a$ and $b$ , you cannot read it at a point (a point has zero width, hence zero sand); you must integrate the density over $[a, b]$ . The total sand is $1$ kg, matching $\int f = 1$ . A tall, narrow spike means sand piled densely (high probability nearby) but covering little width; the area of that spike is what counts. This is exactly why $f (x)$ can exceed $1$ : a Uniform $(0, \frac{1}{2})$ distribution has $f = 2$ on its support — high density over a short interval still integrating to $1$ .

Density curve: probability is the shaded area, not the curve height

Worked Examples

Exponential Lifetime — Mean and a Tail Probability

Given: A component's lifetime

T

(hours) is Exponential with rate

λ = 0.01 hr^{- 1}

, density

f (t) = λ e^{- λ t}, t \geq 0, f (t) = 0, t < 0.

Step 1 — Check normalization.

\int_{0}^{\infty} λ e^{- λ t} d t = [- e^{- λ t}]_{0}^{\infty} = 1.

✓ Step 2 — Mean lifetime via [8]:

μ = \int_{0}^{\infty} t λ e^{- λ t} d t = \frac{1}{λ} = \frac{1}{0.01} .

(Integration by parts:

u = t, d v = λ e^{- λ t} d t

.) Step 3 — Probability it survives past 150 hr using the CDF

F (t) = 1 - e^{- λ t}

P (T > 150) = 1 - F (150) = e^{- 0.01 \cdot 150} = e^{- 1.5} .

μ = 100 hr, P (T > 150) = e^{- 1.5} \approx 0.223

Transforming a Uniform Variable

Given:

X \sim

Uniform

(0, 1)

with

f_{X} (x) = 1

[0, 1]

. Define

Y = - ln X

. Find

f_{Y} (y)

. Step 1 — Invert.

y = - ln x \Rightarrow x = e^{- y}

, valid for

y > 0

(since

x \in (0, 1)

). Step 2 — Jacobian.

\frac{d x}{d y} = - e^{- y}

, so

\frac{d x}{d y} = e^{- y}

. Step 3 — Apply [9].

f_{Y} (y) = f_{X} (e^{- y}) \cdot e^{- y} = 1 \cdot e^{- y} = e^{- y}, y > 0.

Step 4 — Identify. This is Exponential with rate

λ = 1

f_{Y} (y) = e^{- y} (y > 0) ⟹ Y \sim Exponential (1)

This is the basis of inverse-transform sampling: applying

- ln

to uniform draws produces exponential samples.

Common Misconceptions

The density is not a probability.

f (x)

has units of (probability per unit

x

) and may be arbitrarily large. Writing "

P (X = x) = f (x)

" is wrong; the correct statement is

P (X = x) = 0

and

P (x \leq X \leq x + d x) \approx f (x) d x

. Likewise, never claim "

f (x) \leq 1

" — only

\int f = 1

and

f \geq 0

are required. Probability zero ≠ impossible. Every single outcome of a continuous variable has probability

0

, yet exactly one outcome occurs on each trial. "Almost surely" (probability

1

) is not the same as "always." This is why, when conditioning on a continuous variable taking a precise value, one must use densities and limits, not naive

P (A ∣ B) = P (A \cap B) / P (B)

(which gives

0/0

A Continuous CDF With No Density

The Cantor (devil's staircase) distribution has a CDF $F$ that is continuous everywhere and increases from $0$ to $1$ , so $P (X = c) = 0$ for all $c$ — yet $F^{'} (x) = 0$ at almost every point and $F$ is not absolutely continuous. There is no density $f$ with $F (x) = \int_{- \infty}^{x} f$ . Thus "continuous CDF" is strictly weaker than "has a density." A density exists only for absolutely continuous distributions. This shows the chain [3]–[4] can fail without absolute continuity.

Connections

Continuous variables sit opposite Discrete Random Variables, where probability mass functions replace densities and sums replace integrals; both are special cases of distributions described by a Cumulative Distribution Function. The density is rigorously the Radon–Nikodym Derivative of the distribution with respect to Lebesgue measure, tying the subject to Measure Theory. Expectation [7] is an instance of Lebesgue Integration. Two or more continuous variables are handled with Joint Probability Densities and marginals, and independence is characterized by factorization of joint densities. The transformation rule [9] generalizes to the multivariate Jacobian Change of Variables. Specific densities — the Normal Distribution, Exponential Distribution, and Uniform Distribution — are the workhorse examples.

Real-World Applications

Reliability engineering — component lifetimes modeled by Exponential/Weibull densities; the integral [1] gives warranty failure probabilities.
Signal processing & noise — thermal (Johnson–Nyquist) noise voltage follows a Normal Distribution; densities set detection thresholds.
Quantitative finance — the Black–Scholes model assumes log-returns are normal; option prices are expectations [7] over a lognormal density.
Quantum mechanics — $∣ ψ (x) ∣^{2}$ is a genuine probability density for particle position, normalized by [2] (the Born rule).
Statistical inference — maximum-likelihood estimation maximizes products of densities; confidence intervals integrate sampling densities.
Queueing & telecom — inter-arrival and service times use Exponential/Gamma densities to compute waiting-time distributions.
Metrology — measurement error is modeled with Gaussian densities; reported uncertainties are density standard deviations $σ$ from [8].

Historical Development

1733 — Abraham de Moivre derives the bell-curve approximation to the binomial, the first appearance of the normal density.
1774 — Pierre-Simon Laplace introduces continuous error distributions and integrates densities to assess astronomical errors.
1809 — Carl Friedrich Gauss formalizes the normal density in the context of least-squares estimation.
1812 — Laplace's Théorie analytique des probabilités systematizes integral-based probability and characteristic functions.
1837 — Siméon Poisson and contemporaries develop the analytic treatment of continuous limits.
1898 — Émile Borel develops measure theory on the line, the geometric backbone for "area = probability."
1902 — Henri Lebesgue's integral provides the rigorous tool for [7] and unbounded densities.
1913 — Radon (later Nikodym, 1930) prove the Radon–Nikodym theorem, defining the density as a derivative of measures.
1933 — Andrey Kolmogorov's axioms unify discrete and continuous probability under measure theory, the modern standard.

Subtleties & Limits

The density is determined only up to a set of Lebesgue measure zero — two densities differing at countably many points describe the same distribution, so "the" density is really an equivalence class. Distributions can be purely continuous yet singular (Cantor type), having no density at all; the Lebesgue decomposition splits any distribution into absolutely continuous, discrete, and singular parts, and characterizing the singular component of empirical distributions remains subtle. In infinite dimensions (e.g. Gaussian measures on function spaces), there is no Lebesgue reference measure, so densities must be taken relative to another measure (Cameron–Martin theory) — a live area in stochastic analysis. Finally, density estimation from finite data (choosing bandwidths, avoiding the curse of dimensionality) is an active statistical and machine-learning research frontier.