I'm reading a statistics textbook which defines the mean of a random variable $X$ with CDF $F$ as a statistical function $t(\centerdot)$, where

$$ t(F) = \int x \, dF(x).$$

Can someone explain this definition? I'm familiar with the definition of the mean as an integral over the PDF $f$:

$$ \int x \, f(x) \, dx.$$

But what does it mean to have the function $F(x)$ in the variable of integration?


The former integral is a Stieltjes integral. See this, for example, in particular the section "Application to probability theory".

In brief, the integral $I_1 = \int {xdF(x)} $ is a generalization of $I_2 = \int {xf(x)dx} $. $I_2$ can be used only when the distribution is absolutely continuous, that is, has a probability density function. $I_1$ can be used for any distribution, even if it is discrete or continuous singular (provided the expectation is well-defined, that is $\int {|x|dF(x)} < \infty $). Note that if $F'(x) = f(x)$, then $\frac{d}{{dx}}F(x) = f(x)$ gives rise to $dF(x) = f(x)dx$.

For a thorough account of this topic, see, for example, this.


An example. Suppose that a random variable $X$ has CDF $F$ such that $F'(x) = f_1 (x)$ for $x < a$, $F'(x) = f_2 (x)$ for $x > a$, and $F(a) - F(a-) = p$, where $f_1$ and $f_2$ are nonnegative continuous functions, and $0 < p \leq 1$. Note that $F$ is everywhere differentiable except at $x=a$ where it has a jump discontinuity of size $p$. In particular, $X$ does not have a PDF $f$, hence you cannot compute ${\rm E}(X)$ as the integral $I_2$. However, ${\rm E}(X)$ can be computed as follows: $$ {\rm E}(X) = \int_{ - \infty }^\infty {xdF(x)} = \int_{ - \infty }^a {xf_1 (x)dx} + a[F(a) - F(a - )] + \int_a^\infty {xf_2 (x)dx} . $$ In case $f_1$ and $f_2$ are identically zero and, hence, $p=1$, this gives $$ {\rm E}(X) = \int_{ - \infty }^\infty {xdF(x)} = a[F(a) - F(a - )] = a, $$ which is obvious since $X$ has the $\delta_a$ distribution (that is, ${\rm P}(X=a)=1$).

Exercise 1. Suppose that $X$ takes finitely many values $x_1,\ldots,x_n$, with probabilities $p_1,\ldots,p_n$, respectively. Conclude that ${\rm E}(X) = \int_{ - \infty }^\infty {xdF(x)} = \sum\nolimits_{k = 1}^n {x_k p_k } $.

Exercise 2. Suppose that with probability $0 < p < 1$ a random variable $X$ takes the value $a$, and with probability $1-p$ it is uniform$[0,1]$. Find the CDF $F$ of $X$, and compute its expectation (note that $X$ is neither discrete nor continuous random variable).

Note: The integral $I_1$ is, of course, a special case of $\int {h(x)dF(x)} $ (say, for $h$ a continuous function). In particular, letting $h=1$, we have $\int {dF(x)} = 1$ (which is a generalization of $\int {f(x)dx} = 1$).

EDIT: Further details.

Note that if $h$ is continuous, then ${\rm E}[h(X)]$, if exists, is given by $$ {\rm E}[h(X)] = \int {h(x)dF(x)}. $$ In particular, if the $n$-th moment exists, it is given by $$ {\rm E}[X^n] = \int {x^n dF(x)}. $$

In principle, the limits of integration range from $-\infty$ to $\infty$. In this context, consider the following important example. Suppose that $X$ is any nonnegative random variable. Then its Laplace transform is $$ {\rm E}[e^{ - sX} ] = \int {e^{ - sx} dF(x)} = \int_{ 0^- }^\infty {e^{ - sx} dF(x)}, \;\; s \geq 0, $$ where $0^-$ can be replaced by $-\varepsilon$ for any $\varepsilon > 0$. While the $n$-th moment of (the nonnegative) $X$ is given, for any $n \geq 1$, by $$ {\rm E}[X^n] = \int_{ 0^- }^\infty {x^n dF(x)} = \int_0^\infty {x^n dF(x)}, $$ it is not true in general that also $$ {\rm E}[e^{ - sX} ] = \int_{0 }^\infty {e^{ - sx} dF(x)}. $$ Indeed, following the definition of the integral, $$ {\rm E}[e^{ - sX} ] = \int_{ 0^- }^\infty {e^{ - sx} dF(x)} = e^{-s0}[F(0)-F(0-)] + \int_{ 0 }^\infty {e^{ - sx} dF(x)}, $$ hence the jump of $F$ at zero should be added (if positive, of course). In the $n$-th moment case, on the other hand, the corresponding term is $0^n[F(0)-F(0-)] = 0$, hence a jump of $F$ at zero does not affect the overall integral.

| cite | improve this answer | |

The cumulative density function $F$ is related to the probability density function $f$ by $dF(x)/dx=f(x)$. The equation you have in terms of $F$ can be re-expressed in terms of $f$ by substituting in $dF(x) = f(x)\,dx$. In fact, for many purposes, you can take this as the definition of the differential term $dF$. However, in more general circumstances where $F$ is not differentiable and the PDF $f$ is not well-defined, the form involving $dF$ still holds (interpreting it as a Riemann-Stieltjes integral). For example, if the distribution is discrete, so that $F$ is piecewise constant, then $dF$ becomes a sum over Dirac distributions.

| cite | improve this answer | |

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.