Wikipedia presents a definition of the Lebesgue integral (of a nonnegative function $f$) that I hadn't encountered before:

Let $f^*(t)=\mu \left (\{x\mid f(x)>t\} \right )$.
The Lebesgue integral of $f$ is then defined by

$\int f\,d\mu = \int_0^\infty f^*(t)\,dt$
where the integral on the right is an ordinary improper Riemann integral

My question is, what is the relation between this definition and more standard definitions, like the supremum of $\int \phi\,d\mu$ over simple functions $\phi$ such that $0 \leq \phi \leq f$?

I'm also interested in understanding the intuitive justification provided for this definition:

Using the "partitioning the range of $f$" philosophy, the integral of $f$ should be the sum over $t$ of the area of the thin horizontal strip between $y = t$ and $y = t + dt$. This area is just

$\mu \left (\{x\mid f(x)>t\} \right ) \,dt$.

I don't see why that's the area of the infinitesimal strip. Clearly the width of the strip is $dt$, but why is the length of the strip $\mu \left (\{x\mid f(x)>t\} \right ) \,$?

Any help would be greatly appreciated.

Thank You in Advance.

EDIT: It seems to me that the same reasoning that shows that the Lebesgue integral of $f$ can be expressed as $\int_0^\infty f^*(t)\,dt$ can also be used to express $f$ itself as $f(x) = \int_0^\infty f^{**}(x,t)\,dt$, where $f^{**}(x,t) = \chi_{\{s:f(s)>t\}}(x)$. Am I right about that?

  • $\begingroup$ Tell me if I'm wrong, but I think this particular characterization of the Lebesgue integral is mentioned far more often when the subject is probability than when it's anything else. $\endgroup$ – Michael Hardy Apr 23 '14 at 18:01

Say you have a nonnegative function $f$. Let $M > 0$ be fixed and let $n \in \mathbb N$. Partition the range of $f$ into the sets

  • $\left\{\dfrac{kM}{2^n} < f \le \dfrac{(k+1)M}{2^n}\right\}$ for $0 \le k \le 2^n - 1$, and
  • $\{f > M\}$.

Approximate the integral of $f$ by $$ \sum_{k=0}^{2^n - 1} \frac{kM}{2^n} \mu \left( \left\{ \frac{kM}{2^n} < f \le \frac{(k+1)M}{2^n} \right\} \right) + M \mu(\{f > M\}).$$

It isn't hard to see that this expression approximates the integral of $f$ because if $f$ is integrable then $M \mu(\{f > M\}) \to 0$ as $M \to \infty$, and the simple function $$\sum_{k=0}^{ 2^n - 1} \frac{kM}{2^n} \chi_{\{\frac{kM}{2^n} < f \le \frac{(k+1)M}{2^n}\}}(x)$$ approximates $f$ uniformly on the set $\{f \le M\}$.

Define $a_k = \dfrac{kM}{2^n}$ and $b_k = \displaystyle \mu \left( \left\{ f > \frac{kM}{2^n} \right\} \right)$. The sum above may be written as $$\sum_{k=0}^{2^n - 1} a_k (b_k - b_{k+1}) + a_{2^n} b_{2^n}.$$ Now employ the summation-by-parts trick to find this equal to $$ \sum_{k=1}^{2^n} (a_k - a_{k-1}) b_k = \sum_{k=1}^{2^n} \frac{M}{2^n} \mu( \{f > kM/2^n\}).$$As $n \to \infty$, the latter integral converges to $$\int_0^M \mu(\{f > t\}) \, dt.$$ Finally let $M \to \infty$ to get the integral of $f$.

  • $\begingroup$ Thanks. Do you have any thoughts on Wikipedia's intuitive justification of the definition? Why would the area of the infinitesimal horizontal strip between $t$ and $t+dt$ be $\mu \left (\{x\mid f(x)>t\} \right ) \,dt$? $\endgroup$ – Keshav Srinivasan Apr 23 '14 at 18:31

Umberto P.'s proof is quite rigorous, but I thought I'd try to give a slightly more intuitive (but less rigorous) explanation of how the Wikipedia definition follows from more common ways of thinking about the Lebesgue integral.

Let's partition the range $[0,\infty)$ into $y_1,y_2,y_3,...$ (where $y_1=0$), and let's chooser a point $y_i^*$ from each subinterval $[y_i,y_{i+1}]$. Then we can approximate our function $f$ by assuming that all points $x$ which map to the interval $[y_i,y_{i+1}]$ only map to the point $y_i^*$, so that we're approximating $f$ by the simple function $\sum_{i=1}^{\infty} y_i^*\chi_{\{x:y_i\leq f(x)\leq y_{i+1}\}}$, and thus approximating the Lebesgue integral of $f$ by $\sum_{i=1}^{\infty} y_i^*\mu({\{x:y_i\leq f(x)\leq y_{i+1}\}})$.

Now notice that ${\{x:y_i\leq f(x)\leq y_{i+1}\}}$ is the set of points for which $f(x)$ is between $y_i$ and $y_{i+1}$, or to put it another way $f(x)$ is greater than $y_i$ but is not greater than $y_{i+1}$, so we can express the set as ${\{x: f(x) > y_i\}}-{\{x: f(x) > y_{i+1}\}}$. Thus $\mu({\{x:y_i\leq f(x)\leq y_{i+1}\}})=\mu({\{x: f(x) > y_i\}})-\mu({\{x: f(x) > y_{i+1}\}})$, which by definition is equal to $f^*(y_i)-f^*(y_{i+1})$. (Note: I'm playing fast and loose with the distinction between less than and less than or equal to.)

So we can rewrite our approximation of the Lebesgue integral as $\sum_{i=1}^{\infty} y_i^*(f^*(y_i)-f^*(y_{i+1})) = -\sum_{i=1}^{\infty} y_i^*(f^*(y_{i+1})-f^*(y_i))$. Now we can see that in the limit as the width of the intervals $[y_i,y_i+1]$ goes to zero (i.e. the mesh of the partition goes to zero), this expression becomes the Riemann-Stieltjes integral $-\int_0^\infty ydf^{*}(y)$. Since $f^*$ is monotone, it's of bounded variation, so we're allowed to apply integration by parts to it, so we get $-(\lim_{y\rightarrow\infty}(yf^*(y))-0f^*(0)-\int_0^\infty f^*(y)dy)$. I think we can show that if the Lebesgue integral of $f$ exists, then $\lim_{y\rightarrow\infty}(yf^*(y))=\lim_{y\rightarrow\infty}(y\mu({\{x: f(x) > y\}}))=0$.

So all we're left with in the end is $\int_0^\infty f^*(y)dy$, which is the definition in the Wikipedia article. So at least that makes sense to my satisfaction. I still don't understand, however, Wikipedia's intuitive justification for the result, specifically why the area of the infinitesimal horizontal strip between $y=t$ and $y=t+dt$ is $\mu({\{x: f(x) > t\}})dt$.


Forgive my MS Paint, and let $dt>0$ be a tiny change in height. I believe the Wikipedia argument can be intuited as follows. We approximate the integral of $\color{blue}f$ with the area of the $\color{red}{\text{rectangles}}$. enter image description here

$\color{green}{\text{(1)}} = \{x : f(x) > 0\} = [f>0]$. (This is our entire support.) The area of the first rectangle is $\mu(f>0)dt$.

$\color{green}{\text{(2)}} = [f>dt]$. The area of the second collection of rectangles is $\mu(f>dt)dt$

In each case, note that the base of each rectangle at height $t$ is $[f>t]$. In the end we are left with $$\int f \approx \sum_{k=1}^\infty \mu(f>k dt)dt$$ and in the limit $dt\to 0$ we should expect to recover the form as claimed, $\int f = \int_0^\infty \mu(f>t)dt$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.