Characteristic Functions

Adding independent random things is the most common operation in finance — a portfolio is a sum of positions, a monthly return is a sum of daily ones, a desk's PnL is a sum of trades. And it is genuinely awkward. To get the distribution of a sum you must convolve the two distributions: for every possible total, integrate over every way of splitting that total between the two pieces. Do it twice and the algebra is unpleasant. Do it a thousand times and it is hopeless. Characteristic functions make that operation into multiplication.

The trick you already know

Before logarithms, multiplying two seven-digit numbers by hand was a chore. Then someone noticed you could translate each number into its logarithm, add the logs (easy), and translate back. Nothing about multiplication changed. The work moved to a coordinate system where it was trivial.

Characteristic functions do the same thing for probability distributions. Convolution is the hard operation; there is a translation into "frequency space" where convolution becomes ordinary multiplication; and the translation is reversible, so you can always come home with your answer. Same idea as logarithms, one level up.

The definition, one symbol at a time

The characteristic function of a random variable $X$ is

$$\varphi_X(t) = E\!\left[ e^{itX} \right], \qquad t \in \mathbb{R}.$$

Four symbols to name. $\varphi$ (phi) is the function itself. $t$ is a plain real number you get to choose — think of it as a dial, a "frequency". $i$ is the imaginary unit, the number with $i^2 = -1$. And $E[\cdot]$ is the ordinary expected value.

The only unfamiliar piece is $e^{itX}$, and it is more concrete than it looks. Euler's identity says $e^{i\theta} = \cos\theta + i\sin\theta$, which is exactly a unit arrow rotated by the angle $\theta$. So $e^{itX}$ says: take the value $X$ came out as, multiply it by your dial setting $t$, and spin an arrow of length one by that many radians.

In plain English, $\varphi_X(t)$ is the average of all those arrows, one per possible outcome, weighted by probability.

That reading gives you two properties for free. Set $t = 0$: every arrow points the same way (angle zero), so the average arrow has full length and $\varphi_X(0) = 1$. For any other $t$, you are averaging arrows of length one, so the result can never be longer than one: $|\varphi_X(t)| \le 1$. If the outcomes are tightly bunched, small dials leave the arrows nearly aligned and $|\varphi|$ stays near 1; if the outcomes are spread out, even a modest dial fans the arrows around the circle and they cancel.

The property that does the work

If $X$ and $Y$ are independent, then

$$\varphi_{X+Y}(t) = \varphi_X(t)\,\varphi_Y(t).$$

The characteristic function of a sum is the product of the characteristic functions. The reason is one line: $E[e^{it(X+Y)}] = E[e^{itX} e^{itY}]$, and independence lets the expectation of a product split into the product of expectations. Convolution on the hard side, a single multiplication on the easy side.

The other property that makes this useful is uniqueness: two random variables with the same characteristic function have the same distribution, full stop. So if you multiply two characteristic functions and recognise the answer, you have identified the sum's distribution without ever touching an integral.

Worked example 1: adding two normal returns

A strategy's annual return $X$ is normal with mean $0.05$ and standard deviation $0.20$. An independent strategy $Y$ is normal with mean $0.03$ and standard deviation $0.15$. What is the distribution of the combined return $X + Y$?

The characteristic function of a normal with mean $\mu$ and variance $\sigma^2$ is

$$\varphi(t) = \exp\!\left( i\mu t - \tfrac{1}{2}\sigma^2 t^2 \right).$$

Fill in the two variances: $0.20^2 = 0.04$ and $0.15^2 = 0.0225$. So

$$\varphi_X(t) = \exp\!\left( 0.05\,i t - 0.02\,t^2 \right), \qquad \varphi_Y(t) = \exp\!\left( 0.03\,i t - 0.01125\,t^2 \right).$$

Multiply them. Multiplying exponentials means adding exponents:

$$\varphi_{X+Y}(t) = \exp\!\left( (0.05 + 0.03)\,i t - (0.02 + 0.01125)\,t^2 \right) = \exp\!\left( 0.08\,it - 0.03125\,t^2 \right).$$

Compare with the template: the mean is $0.08$ and $\tfrac{1}{2}\sigma^2 = 0.03125$, so $\sigma^2 = 0.0625$ and $\sigma = \sqrt{0.0625} = 0.25$.

The sum is normal with mean $0.08$ and standard deviation $0.25$. Notice what did not happen: the standard deviations did not add, since $0.20 + 0.15 = 0.35$, not $0.25$. Variances add; volatilities do not. The characteristic function makes that visible in one line, because the variance sits in the exponent where addition happens.

Drag the standard deviation slider below and watch the density widen. The characteristic function moves the opposite way — a wider distribution in $x$ gives a faster-decaying $\varphi$ in $t$, because spread-out outcomes fan the arrows apart sooner.

Worked example 2: reading moments off the curve

Characteristic functions also store the moments. Differentiate at $t = 0$:

$$E[X] = \frac{\varphi'(0)}{i}, \qquad E[X^2] = -\,\varphi''(0).$$

Each derivative at zero pulls out one more moment, with factors of $i$ to keep track of. Test it on a fair coin scored as $X = 1$ for heads and $X = 0$ for tails.

Step 1, build $\varphi$ from the definition. Two outcomes, each with probability $0.5$:

$$\varphi(t) = 0.5\,e^{it \cdot 0} + 0.5\,e^{it \cdot 1} = 0.5 + 0.5\,e^{it}.$$

Step 2, check the sanity condition: $\varphi(0) = 0.5 + 0.5 = 1$. ✓

Step 3, differentiate once. The derivative of $e^{it}$ with respect to $t$ is $i e^{it}$, so $\varphi'(t) = 0.5\,i\,e^{it}$ and $\varphi'(0) = 0.5\,i$. Then $E[X] = 0.5i / i = 0.5$. ✓ — a fair coin scores half a point on average.

Step 4, differentiate again: $\varphi''(t) = 0.5\,i^2 e^{it} = -0.5\,e^{it}$, so $\varphi''(0) = -0.5$ and $E[X^2] = 0.5$. That is right, because $X$ is 0 or 1 so $X^2 = X$.

Step 5, assemble the variance: $\operatorname{Var}(X) = E[X^2] - (E[X])^2 = 0.5 - 0.25 = 0.25$, giving a standard deviation of $0.5$. ✓ Everything matches the answers you would get directly, but obtained by differentiating a single function.

Worked example 3: the distribution that laughs at averaging

Here is the payoff for using characteristic functions instead of moment generating functions. The Cauchy distribution has such fat tails that its mean does not exist and its moment generating function is infinite for every $t \ne 0$. Its characteristic function exists and is beautifully simple:

$$\varphi(t) = e^{-|t|}.$$

Take $n$ independent Cauchy draws and average them: $M_n = \tfrac{1}{n}(X_1 + \cdots + X_n)$. Scaling a variable by $1/n$ replaces $t$ with $t/n$ inside $\varphi$, and independence turns the sum into a product, so

$$\varphi_{M_n}(t) = \left[ \varphi\!\left( \tfrac{t}{n} \right) \right]^{n} = \left[ e^{-|t|/n} \right]^{n} = e^{-|t|} = \varphi(t).$$

Three symbols of algebra and the conclusion is startling: the average of a million Cauchy draws has exactly the same distribution as a single draw. Averaging buys you nothing at all. The law of large numbers fails, error bars never shrink, and no amount of data helps. This is not an exotic curiosity — it is the clean warning about what fat tails do to estimation.

What this means in practice

• Proving the central limit theorem. Expand $\varphi$ for a standardised sum, multiply $n$ copies, and the expression collapses to $e^{-t^2/2}$ — the characteristic function of a standard normal. Lévy's continuity theorem turns that pointwise limit into convergence in distribution. That is the whole proof of the CLT, in about six lines.
• Pricing under jumps and stochastic vol. Heston, Variance-Gamma and Merton jump-diffusion have no closed-form density but a clean closed-form characteristic function. Carr-Madan and COS pricing transform the payoff, multiply, and invert numerically — the diagram above, run on a computer. See Heston and jump diffusion.
• Building distributions from parts. Compound Poisson losses, sums of factors and mixture models are all assembled by manipulating characteristic functions and inverting once at the end.
• Fat-tail diagnostics. How fast $|\varphi(t)|$ decays near zero encodes tail behaviour; a kink at the origin, as in $e^{-|t|}$, signals tails so heavy the mean is undefined.