Definite and Indefinite Integrals

An indefinite integral is the whole family of curves sharing one slope pattern; a definite integral picks two endpoints and returns one number.

Three parallel cubic curves differing by vertical shifts stand for one indefinite integral family, while a shaded band under x squared from 0 to 2 shows the definite integral collapsing to the single number 8 thirds

Why Does This Exist?

Expected values, marginal probabilities, and cumulative metrics are definite integrals: single numbers with units. Recovering a position from velocity data, or a loss from its gradient field, is an indefinite integral: a family of curves differing by a starting offset. Confusing the two produces answers with the wrong shape entirely, a function where a number belongs or a number where a curve belongs.

Integration and area builds the area intuition. This page draws the boundary both notations share: when the integral sign carries limits it computes, and when it does not it describes.

The one idea you need first

An antiderivative of $f$ is any $F$ with $F' = f$ . Since constants differentiate to zero, antiderivatives come in families: if $x^3/3$ works for $x^2$ , so does $x^3/3 + 5$ . That $+C$ is not decoration; it is the whole difference between the two integral kinds.

Think of It Like This

A speedometer with and without a trip meter

An indefinite integral is a speedometer history with no starting odometer: from speeds alone you reconstruct the route's shape but not its position, every reconstruction shifted by the unknown starting mile. The $+C$ is that unknown start.

A definite integral sets two trip-meter presses, at mile 0 and mile 2, and reports one number: distance covered between presses. Same speeds, but the endpoints pin the answer down to $8/3$ miles. Limits are the trip meter; without them you own a shape, with them you own a measurement.

How It Actually Works

Indefinite integral: a function family. $\int f(x)\,dx = F(x) + C$ collects every antiderivative. Example: $\int x^2\,dx = x^3/3 + C$ , because differentiating $x^3/3 + C$ returns $x^2$ for any constant $C$ . No limits, no number, just the family.

Definite integral: one number. $\int_a^b f(x)\,dx = F(b) - F(a)$ evaluates the family at both endpoints and subtracts, so every $C$ cancels: $(F(b) + C) - (F(a) + C) = F(b) - F(a)$ . Limits, one number, units of $f$ times units of $x$ .

Worked example

Integrate $f(x) = x^2$ both ways.

Indefinite: $\int x^2\,dx = x^3/3 + C$ . Verify by differentiating: $\frac{d}{dx}(x^3/3 + C) = x^2$ . Members include $x^3/3$ , $x^3/3 + 1$ , $x^3/3 - 7$ : parallel curves, identical slopes everywhere.

Definite from 0 to 2: $\int_0^2 x^2\,dx = [x^3/3]_0^2 = 8/3 - 0 = 8/3 \approx 2.667$ . One number, about 2.67 square units. Any $C$ you carried cancels in the subtraction, which is why definite answers never mention it.

Watch Out For

Dropping the plus C in the indefinite case

The symptom is a "general solution" that misses most solutions: reporting $x^3/3$ as the antiderivative silently discards every vertical shift, and downstream initial conditions then have nothing to grip. In ML this bites when reconstructing potentials or integrating gradient fields: one integration constant per component, each needing its own condition. Write $+C$ every time the limits are absent, and trade it for a fitted constant the moment data arrives.

Handing a function where a number belongs

The symptom is a probability or expectation left as an unevaluated antiderivative: correct shape, wrong type, and every consumer breaks. Definite integrals answer "how much"; indefinite integrals answer "what shape." If the question names endpoints ("from 0 to 2," "over the batch," "expected value"), evaluate at both and subtract. If it asks for the family ("solve," "recover," "find all"), keep $C$ and stop.

The Quick Version

Indefinite integrals return curve families: $\int x^2\,dx = x^3/3 + C$ .
Definite integrals return numbers: $\int_0^2 x^2\,dx = 8/3 \approx 2.667$ .
The $+C$ cancels in $F(b) - F(a)$ , so definite answers never carry it.
No limits means describe the family; limits mean compute the measurement.
Always verify by differentiating the antiderivative back.
Shape questions want $+C$ ; "how much" questions want endpoints evaluated.