The Lebesgue Integral
The Lebesgue integral adds up a function by slicing its range —
the output values — rather than its domain. A
Riemann integral chops the input
axis into thin vertical strips; the Lebesgue integral instead chops the value axis into
bands and asks, for each band, how much of the space lands that high. Measured
against a
measure
\mu, we write it
\int_{\Omega} f \, d\mu.
Slicing by value is exactly what lets the integral handle badly-behaved
measurable functions
and — crucially — pass cleanly through limits, the convergence theorems the rest of the
theory leans on. It is built in four deliberate stages, each extending the
last: first the indicator of a single set, then a finite combination of indicators (a
simple function), then any non-negative measurable function, and finally a general
signed one. We take them in order and skip no step.
Stage 0 — the indicator of a set
The atom of the whole construction is the indicator function of a
measurable set A \in \mathcal{F}: the function that is
1 on A and 0
off it,
\mathbf{1}_A(\omega) \;=\; \begin{cases} 1 & \omega \in A,\\[2pt] 0 & \omega \notin A. \end{cases}
Its integral is defined to be the measure of the set it marks out — there is nothing
to compute, this is the seed everything else grows from:
\int_{\Omega} \mathbf{1}_A \, d\mu \;=\; \mu(A).
Read it as: "weight 1 over the region A,
weight 0 everywhere else, and total it up" — and the total weight
is just the size \mu(A) of A. With
\mu = \mathbb{P} a probability measure this already says
\int \mathbf{1}_A \, d\mathbb{P} = \mathbb{P}(A), the fact that
makes probability a special case of integration.
Stage 1 — simple functions
A simple function takes finitely many values
a_1, \dots, a_n, each on a measurable set
A_i, where the A_i partition
\Omega (they are disjoint and cover everything). It is just a
finite stack of indicators, one per value:
S \;=\; \sum_{i=1}^{n} a_i \, \mathbf{1}_{A_i}.
Its integral is forced on us by Stage 0 and the demand that integration be additive: integrate
the stack term by term, and each term a_i \mathbf{1}_{A_i}
contributes a_i \mu(A_i). So the integral is the value-weighted
total of the measures of those sets,
\int_{\Omega} S \, d\mu \;=\; \sum_{i=1}^{n} a_i \, \mu(A_i).
In particular, taking n = 1 with
S = \mathbf{1}_A recovers Stage 0,
\int_{\Omega} \mathbf{1}_A \, d\mu = \mu(A).
Linearity on simple functions, line by line
The one fact the rest of the theory rests on is that
S \mapsto \int S \, d\mu is linear. It is worth
seeing exactly why, because the only subtlety is that two simple functions are written over
different partitions, and we must put them on a common one before we can add.
Take two simple functions over their own partitions of \Omega,
S \;=\; \sum_{i=1}^{m} a_i \, \mathbf{1}_{A_i}, \qquad T \;=\; \sum_{j=1}^{n} b_j \, \mathbf{1}_{B_j}.
Form the common refinement: the sets
C_{ij} = A_i \cap B_j. Because
\{A_i\} and \{B_j\} each partition
\Omega, the C_{ij} are disjoint and also
partition \Omega; and crucially, for a fixed
i the set A_i is sliced exactly into its
pieces C_{ij}, so by finite additivity of the measure
\mu,
\mu(A_i) \;=\; \sum_{j=1}^{n} \mu(C_{ij}), \qquad \mu(B_j) \;=\; \sum_{i=1}^{m} \mu(C_{ij}).
On each tiny cell C_{ij} both functions are constant —
S = a_i and T = b_j there — so for any
scalars \alpha, \beta the combination
\alpha S + \beta T is itself simple, equal to
\alpha a_i + \beta b_j on C_{ij}:
\alpha S + \beta T \;=\; \sum_{i=1}^{m}\sum_{j=1}^{n} (\alpha a_i + \beta b_j)\, \mathbf{1}_{C_{ij}}.
Now apply the Stage 1 definition to this simple function and unpack it step by step:
\int (\alpha S + \beta T)\, d\mu \;=\; \sum_{i=1}^{m}\sum_{j=1}^{n} (\alpha a_i + \beta b_j)\, \mu(C_{ij})
by the definition of the integral of a simple function. Split the sum across the plus sign:
=\; \alpha \sum_{i=1}^{m}\sum_{j=1}^{n} a_i\, \mu(C_{ij}) \;+\; \beta \sum_{i=1}^{m}\sum_{j=1}^{n} b_j\, \mu(C_{ij})
by regrouping the finite double sum. In the first double sum a_i
does not depend on j, so sum over j first;
in the second b_j does not depend on i, so
sum over i first:
=\; \alpha \sum_{i=1}^{m} a_i \Big(\sum_{j=1}^{n}\mu(C_{ij})\Big) \;+\; \beta \sum_{j=1}^{n} b_j \Big(\sum_{i=1}^{m}\mu(C_{ij})\Big)
and the two bracketed sums are exactly the additivity identities above, so they collapse:
=\; \alpha \sum_{i=1}^{m} a_i\, \mu(A_i) \;+\; \beta \sum_{j=1}^{n} b_j\, \mu(B_j) \;=\; \alpha \int S \, d\mu \;+\; \beta \int T \, d\mu.
That is linearity, with no gaps: the common refinement is the whole trick, and additivity of
\mu does the rest. Monotonicity is even quicker — if
S \le T, then on every cell
a_i \le b_j, and since each \mu(C_{ij}) \ge 0
the term-by-term inequality survives summation, giving
\int S \, d\mu \le \int T \, d\mu.
Stage 2 — non-negative functions
For a general measurable f \ge 0 we cannot write down a finite
formula, so instead we approximate it from below by simple functions and take the
best such approximation — the supremum of their integrals:
\int_{\Omega} f \, d\mu \;=\; \sup\Big\{\, \textstyle\int_{\Omega} S \, d\mu \;:\; S \text{ simple},\; 0 \le S \le f \,\Big\}.
Two things make this definition sound. First, the set of candidates is non-empty — the simple
function S \equiv 0 always qualifies — so the supremum is over a
non-empty set of real numbers. Second, by monotonicity on simple functions
(just proved), every candidate already satisfies
\int S \, d\mu \le \int S' \, d\mu whenever
S \le S', so raising the staircase can only raise the estimate — the
supremum genuinely captures the "best from below".
The trick is to partition the range, not the domain: slice the
y-axis into finer and finer levels and, on each level, ask "how
much of the space reaches this high?". Concretely, with n levels of
height h = \tfrac{1}{n} one can take the simple function that floors
f down to the nearest level below it; as n
grows the staircase climbs up to f and the simple-function integrals
increase to \int f \, d\mu. The interactive figure below shows
exactly this staircase tightening against the curve.
Let 0 \le f_1 \le f_2 \le \cdots be a non-decreasing sequence of
non-negative measurable functions with pointwise limit
f = \lim_{n\to\infty} f_n. Then the limit and the integral may be
exchanged:
\int_{\Omega} \Big(\lim_{n\to\infty} f_n\Big)\, d\mu \;=\; \lim_{n\to\infty} \int_{\Omega} f_n \, d\mu.
This is precisely what licenses the Stage 2 definition: the flooring staircases
S_n \uparrow f are such a non-decreasing sequence, so their integrals
rise to \int f \, d\mu with no loss in the limit.
Stage 3 — general functions
A general f takes both signs, but Stage 2 only knows how to
integrate non-negative functions. So split f into its
positive and negative parts,
f^{+} \;=\; \max(f, 0), \qquad f^{-} \;=\; \max(-f, 0),
each of which is non-negative and measurable. They reconstruct
f and its absolute value cleanly:
f \;=\; f^{+} - f^{-}, \qquad |f| \;=\; f^{+} + f^{-},
and at every point \omega exactly one of
f^{+}(\omega), f^{-}(\omega) is non-zero — the positive and negative
parts are disjointly supported. Each part is non-negative, so each has a Stage 2
integral; define
\int_{\Omega} f \, d\mu \;=\; \int_{\Omega} f^{+} \, d\mu \;-\; \int_{\Omega} f^{-} \, d\mu,
provided the two pieces are not both +\infty (otherwise the
difference is the meaningless \infty - \infty). We call
f integrable precisely when both pieces are finite,
which by |f| = f^{+} + f^{-} and Stage 1 linearity is the single
condition
\int_{\Omega} |f| \, d\mu \;=\; \int_{\Omega} f^{+} \, d\mu + \int_{\Omega} f^{-} \, d\mu \;<\; \infty.
Two properties now extend from the simple-function case to all integrable
f, g — by approximating each part from below and passing to the limit
with Monotone Convergence — and they are used on essentially every page that follows:
- Linearity: \int (a f + b g) \, d\mu = a\!\int f \, d\mu + b\!\int g \, d\mu.
- Monotonicity: if f \le g then \int f \, d\mu \le \int g \, d\mu.
The picture that separates the two integrals is which axis you cut. The
Riemann integral chops the
domain into thin vertical strips and sums
f(x_i)\,\Delta x_i — it asks "where is
x?". The Lebesgue integral chops the range into
horizontal bands and, for each band height y, weighs the set
\{\,\omega : f(\omega) \ge y\,\} by \mu —
it asks "how much of the space reaches this high?". Slicing the range is what frees the
integral from needing a tidy domain: the indicator of the rationals,
\mathbf{1}_{\mathbb{Q}}, has no Riemann integral (its strips never
settle), yet its Lebesgue integral is plainly \mu(\mathbb{Q}) = 0.
Slicing the range is also why limits pass through so cleanly — the single most
important advantage for probability, where we constantly take limits of approximating random
variables and of conditional estimates. Monotone Convergence handles increasing sequences;
its dominated cousin handles the general case:
Dominated Convergence. If f_n \to f pointwise and
all the f_n are bounded by one fixed integrable
g (so |f_n| \le g with
\int g \, d\mu < \infty), then
\lim_{n\to\infty} \int_{\Omega} f_n \, d\mu \;=\; \int_{\Omega} \lim_{n\to\infty} f_n \, d\mu \;=\; \int_{\Omega} f \, d\mu.
The dominating g is the safety net that stops mass from leaking off
to infinity. This is the theorem that lets us swap a limit and an expectation — exactly the
move behind continuity of expectation, the tower property of conditioning, and the
interchange of differentiation and integration that pricing formulas rely on.
The integral of an indicator is the measure of the set, not automatically
1:
\int \mathbf{1}_A\,d\mu = \mu(A). And simple functions add by
value-weighted measure,
\int \sum_i a_i\mathbf{1}_{A_i}\,d\mu = \sum_i a_i\mu(A_i) —
you cannot drop the measures or just sum the heights. Linearity is with respect to the
integrand, not a licence to ignore the underlying measure.