Let $(\Omega, \mathcal{A}, P)$ be a probability space and $(\Omega', \mathcal{A}')$ be a measurable space. Furthermore, let $X :\Omega \rightarrow \Omega'$ be a random variable.

If $(\Omega', \mathcal{A}') = (\mathbb{R}, \mathcal{B}_{\mathbb{R}})$, then the expected value of $X$ is defined (if it exists) as the Lebesgue integral of $X$ over the measure space $(\Omega, \mathcal{A}, P)$ :

$$ E(X):= \int_{\Omega} X \ dP. $$

I am currently learning about characteristic functions of random variables. The usual definition for the characteristic function $\phi_X : \mathbb{R} \rightarrow \mathbb{R}$ of a random variable is:

$$ \phi(t) := E(e^{itX}) .$$

Usually, no further explanation is given. I have problems bringing this definition together with the definition of the expected value as a Lebesgue integral.

1.) The function $e^{itX}$ is complex-valued and the Lebesgue-integral is only defined for real-valued functions. How can this be resolved?

2.) Is there a general way to define a Lebesgue-integral for complex-valued measurable functions $g : \Omega \rightarrow \mathbb{C}$. If yes, how does it work and what is the usual $\sigma$-algebra on $\mathbb{C}$?

I am happy to hear any thoughts on this to get some more clarity about this obscure definition. I am interested to understand the definition of a characteristic function in the most rigorous way possible. Thanks in advance for any help.


For a given complex-valued function $f;\Omega \to \mathbb{C}$ we can use the decomposition

$$f(x) = \text{Re} \, f(x) + i \, \text{Im} \, f(x)$$

to define the integral of $f$. Indeed, since the real part and the imaginary part are both real-valued functions, we know how to integrate them, and therefore the following definition makes sense:

$$\int f(x) \, \mu(dx) := \int \text{Re} f(x) \, \mu(dx) + i \int \text{Im} \, f(x) \, \mu(dx) \tag{1}$$

where $\mu$ is some measure on the measurable space $(\Omega,\mathcal{A})$ (e.g. the Lebesgue measure or the distribution of a random variable). Obviously, we need to ensure that the integrals on the right-hand side of $(1)$ are well-defined, i.e. we need $\text{Re} f \in L^1(\mu)$ and $\text{Im} \, f \in L^1(\mu)$.

Note that the so-defined integral is linear - this is one of the most important properties of the integral. Moreover, if $f$ is a real-valued function, then we recover the "classical" integral for real-valued functions.

The most common $\sigma$-algebra on $\mathbb{C}$ is the Borel-$\sigma$-algebra $\mathcal{B}(\mathbb{C})$, i.e. the $\sigma$-algebra induced by the metric

$$d(z_1,z_2) := |z_1-z_2| := \sqrt{(\text{Re}(z_1-z_2))^2 + (\text{Im}(z_1-z_2))^2}.$$

It is possible to show that a function $f: (\Omega,\mathcal{A}) \to (\mathbb{C},\mathcal{B}(\mathbb{C}))$ is measurable if, and only if, $\text{Re}\, f: (\Omega,\mathcal{A}) \to (\mathbb{R},\mathcal{B}(\mathbb{R}))$ and $\text{Im} \, f: (\Omega,\mathcal{A}) \to (\mathbb{R},\mathcal{B}(\mathbb{R}))$ are measurable. This stems from the fact that the bijective mapping

$$\iota: \mathbb{C} \to \mathbb{R}^2, z \mapsto (\text{Re} \, z, \text{Im} \, z)$$

is bi-measurable, i.e. both $\iota$ and its inverse are (Borel-)measurable.


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.