Step 04 · Spectral & Models
Spectral Density Notes
Every stationary record can be read a second way: not as a line moving through time but as a spread of variance across frequencies. The spectral density is that portrait, and it often reveals more than the original line.
Variance as a resource
The total variance of a record is fixed and finite. The central insight of the spectral view is that this fixed quantity can be divided among the many oscillations of which the record is woven. A slow oscillation, one that completes few cycles across the record, receives a share; a fast oscillation, completing many cycles, receives another. The spectral density reports the size of each share as a function of frequency, and by construction the shares sum back to the total variance.
Matching shapes to records
A record that wanders slowly, drifting high for long stretches and low for long stretches, carries its variance in the slow end of the range, and its spectral density shows a mound there. A record that flips sharply from one value to another carries its variance in the fast end, and the density shows a mound at the far right. A record with a regular rhythm, such as a seasonal swing, shows a sharp spike at the frequency of that rhythm. Reading the density, then, is reading a summary of the record's behaviour in one glance.
The bridge to autocorrelation
The spectral density and the autocorrelation function carry the same information in two different costumes. They are linked by a transform that turns the autocorrelation sequence into the density and back. A slowly decaying autocorrelation means a density crowded at low frequency; a rapidly decaying autocorrelation means a density spread broadly across the range. Acknowledging this pairing is useful because it lets a reader choose whichever view is easier for the question at hand.
Peaks as rhythms
When a spectral density shows a sharp isolated peak, the record contains a rhythm at that frequency. When it shows a broad plateau, the record contains noise spread evenly across frequencies. Most real records show a mixture: a broad base of noise with one or two brighter peaks standing above it. Separating base from peaks is a central labour of the spectral view.
Why frequency language helps
Time is not always the easiest axis for a question. Ask how a record responds to a gradual change and the low-frequency share answers directly. Ask how it responds to a sudden jolt and the high-frequency share answers. By shifting attention from moments to frequencies, the spectral view gives every question a natural home, and it explains why two records that look equally restless in a plot can behave quite differently when their shares are told apart.
A worked contrast in one paragraph
Take two records of the same length and the same total variance. The first is a slow drift that spends long stretches above its centre and long stretches below. The second is a restless flicker that turns at every step. In a plot the two may look equally busy, yet their spectral portraits could hardly differ. The drift carries its variance in the slowest shares, so its density mounds at the left. The flicker carries its variance in the fastest shares, so its density mounds at the right. The same total, divided into two opposite shapes.
This contrast is the reason the spectral view earns its place in the compendium. It separates two records that a single summary would call identical, and it does so without inventing any information the record did not already contain. The division of variance was always there; the spectral view simply makes it visible.