The spectral density is a beautiful object — but it belongs to the process, an idealised generating mechanism we never actually see. In practice you have one finite stretch of data, x_1, \dots, x_T, and a question: how do I estimate the spectrum from that? The classical answer, invented by Arthur Schuster in the 1890s to hunt for hidden cycles in sunspot and weather records, is the periodogram — and it comes with a famous, brutal catch that every practitioner has to learn.

Schuster's idea is direct: project the data onto sines and cosines at each frequency and measure how much "sticks". A big response at some frequency means the data lean strongly on that rhythm.

The definition

We evaluate the estimate at the Fourier frequencies \omega_j = 2\pi j / T for j = 1, \dots, \lfloor T/2\rfloor — the natural grid of frequencies that fit a whole number of cycles into the sample. The periodogram is

Inside the bars is a Fourier series coefficient of the data; squaring its magnitude turns it into a power. There is an equivalent real form that makes the projection explicit — with a_j = \tfrac{2}{T}\sum_t x_t\cos(\omega_j t) and b_j = \tfrac{2}{T}\sum_t x_t\sin(\omega_j t), I(\omega_j) = \tfrac{T}{4}(a_j^2 + b_j^2). A pure sinusoid buried in the data lights up its own Fourier frequency like a beacon.

Spiky raw, smooth truth

Below, the jagged line is a raw periodogram of a series whose true spectrum has a single hump. The bold line is the smoothed estimate — the same ordinates averaged over a small window of neighbouring frequencies. The hump at the true frequency survives the smoothing; the surrounding grass gets mowed flat.

The raw peak is real, but so is all the surrounding spikiness — and none of that spikiness would go away if you collected ten times as much data. Only averaging removes it.

Fixing an estimator that won't sit still

Because more data does not tame the variance, we must trade a little resolution for a lot of stability. Three standard moves, all forms of averaging:

• Smoothing. Replace I(\omega_j) with a weighted average of it and its neighbours (a spectral window, e.g. the Daniell kernel). Averaging m nearly-independent ordinates cuts the variance by about 1/m at the cost of blurring fine detail — the classic bias–variance trade-off.
• Tapering. Multiply the data by a bell-shaped window before transforming, so the ends fade smoothly to zero. This suppresses leakage — power bleeding from a strong peak into distant frequencies through the sharp edges of a finite record.
• Welch's method. Split the record into overlapping segments, periodogram each, and average the results — smoothing across segments rather than across frequencies.

Each buys consistency with a controlled loss of resolution. A smoothed spectral estimate does converge to f(\omega) as the sample grows and the window narrows at the right rate — the raw periodogram never does.

A worked number

Take T = 100 observations sampled once per month. The Fourier frequency \omega_j = 2\pi j/T that corresponds to an annual (12-month) cycle needs 2\pi j/100 = 2\pi/12, i.e. j = 100/12 \approx 8.3. Since j must be an integer, the nearest Fourier frequencies are j = 8 (period 100/8 = 12.5 months) and j = 9 (period \approx 11.1 months). A true 12-month cycle therefore splits its power between those two bins — an artefact called scalloping that a slightly longer record (say T = 96 = 8\times 12, which puts a Fourier frequency exactly on the annual cycle) would avoid.