Reading Cycles and Persistence from Spectra

What a Spectrum Says About Economic Variation

Economic time series can vary in ways that are difficult to distinguish from a plot alone. Output may move persistently over many quarters, an interest-rate spread may fluctuate in a cyclical pattern, or exchange-rate changes may show little concentration at any particular frequency. A spectrum describes how the variance of a stationary process is distributed across frequencies. Reading its shape helps connect the observed timescale of variation to the process’s dependence over time.

The key is to interpret the spectrum and autocovariance together. A high value near frequency zero is associated with persistent, slowly changing movements. A concentration around a nonzero frequency is associated with cyclical movements and oscillating autocovariances. Power spread broadly across frequencies is less concentrated at a particular timescale. These are descriptions of second-order behavior; by themselves, they do not identify a causal mechanism.

Definitions and Frequency Interpretation

Definition: Let $X_t$ be a real-valued covariance-stationary process with mean $\mu$, and let $Y_t=X_t-\mu$. Write $\gamma(h)=\operatorname{Cov}(Y_t,Y_{t-h})$. For this tutorial, assume the process has a spectral density and, specifically, that $\sum_{h=-\infty}^{\infty}|\gamma(h)|<\infty$. Under the convention used here, $$ f(\omega)=\frac{1}{2\pi}\sum_{h=-\infty}^{\infty}\gamma(h)e^{-ih\omega},\qquad -\pi\leq\omega\leq\pi. $$ The inverse relation is $\gamma(h)=\int_{-\pi}^{\pi} f(\omega)e^{ih\omega}\,d\omega$. Thus $\gamma(0)=\operatorname{Var}(X_t)=\int_{-\pi}^{\pi}f(\omega)\,d\omega$.

A frequency $\omega$ is measured in radians per observation. For $0<\omega\leq\pi$, its associated period is $2\pi/\omega$ observations per cycle: a frequency of $\pi/4$, for example, corresponds to a period of eight observations. Frequency zero corresponds to variation that changes very slowly relative to the observation interval. The spectrum of a real-valued process is nonnegative, periodic, and symmetric around zero, so a cyclical feature at a positive frequency has a matching feature at its negative.

The spectral density’s units are the squared units of $X_t$ per radian of frequency. A density value at one frequency is not the variance “at that exact frequency”; variance over a frequency band is the integral of $f(\omega)$ over that band. A narrow and tall feature can therefore indicate concentration in a small frequency range without implying that the process has greater total variance.

Low-Frequency Concentration and Persistence

When autocovariances are positive across many lags, they add constructively near zero frequency. Under the absolute-summability condition above, symmetry of the autocovariance gives

f(0)=\frac{1}{2\pi}\sum_{h=-\infty}^{\infty}\gamma(h) =\frac{1}{2\pi}\left(\gamma(0)+2\sum_{h=1}^{\infty}\gamma(h)\right).

This identity explains why sustained positive dependence often produces a high spectrum near zero. Shocks tend to have effects that remain aligned over several observations, so relatively slow movements account for substantial variation. The implication is not automatic for every process: negative or oscillating autocovariances can cancel in the sum, and the size of $f(0)$ also depends on the scale of the process.

Worked Example: Persistent Output Deviations

Suppose a hypothetical stationary output-deviation process follows $Y_t=0.8Y_{t-1}+\varepsilon_t$, where the innovations are uncorrelated over time, have mean zero and variance $1$, and the process is the stationary solution. Since the coefficient has absolute value less than one, its autocovariances are absolutely summable. The spectral density is

f(\omega)=\frac{1}{2\pi}\frac{1}{1+0.8^2-2(0.8)\cos(\omega)}.

At zero frequency, the denominator is $1+0.64-1.6=0.04$, so

f(0)=\frac{1}{2\pi(0.04)}=3.979.

At the highest frequency, $\omega=\pi$, the denominator is $1+0.64+1.6=3.24$, giving

f(\pi)=\frac{1}{2\pi(3.24)}=0.04909,\qquad \frac{f(0)}{f(\pi)}=\frac{3.24}{0.04}=81.

The spectrum is strongly concentrated at low frequencies relative to its value at $\pi$. The autocovariance is $\gamma(h)=0.8^{|h|}/(1-0.8^2)$, so the variance is $1/0.36=2.778$, and the lag-five autocorrelation is $0.8^5=0.3277$. Both descriptions point to persistent deviations: the spectrum emphasizes slow movements, while the autocorrelation shows that dependence decays over time rather than disappearing immediately.

Cyclical Peaks and Oscillating Dependence

A spectral concentration around a nonzero frequency indicates that variation at the corresponding timescale is relatively prominent. The associated autocovariances tend to alternate in sign as the lag increases: observations separated by about half a cycle tend to move in opposite directions, while those separated by about a full cycle tend to move in the same direction. The size of the autocovariances also matters. If their oscillations decay slowly, cyclical dependence remains noticeable over more lags.

The following constructed spectrum gives a precise illustration. Define the Poisson kernel $A_r(x)=(1-r^2)/(1-2r\cos x+r^2)$, where $0<r<1$. A symmetric pair of such components produces a real, nonnegative spectral density. Its components are centered at opposite frequencies, as required for a real-valued process.

Worked Example: A Damped Business-Cycle Pattern

Consider a hypothetical stationary economic indicator with spectral density

f(\omega)=\frac{\sigma^2}{4\pi}\left[A_r(\omega-\omega_0)+A_r(\omega+\omega_0)\right], \qquad r=0.95,\quad \omega_0=\frac{\pi}{4}.

The two components are centered at $\omega=\pm\pi/4$, corresponding to a period of $2\pi/(\pi/4)=8$ observations. Since $r$ is close to one, each component is concentrated in a relatively narrow frequency range. The parameter $r$ also controls the decay of the covariance envelope.

The Fourier expansion $A_r(x)=\sum_{h=-\infty}^{\infty}r^{|h|}e^{-ihx}$ gives the autocovariance

\gamma(h)=\sigma^2r^{|h|}\cos(h\omega_0).

For example, the first, second, third, and fourth lag covariances are

\begin{aligned} \gamma(1)&=\sigma^2(0.95)\cos(\pi/4)=0.6718\sigma^2,\\ \gamma(2)&=\sigma^2(0.95)^2\cos(\pi/2)=0,\\ \gamma(3)&=\sigma^2(0.95)^3\cos(3\pi/4)=-0.6063\sigma^2,\\ \gamma(4)&=\sigma^2(0.95)^4\cos(\pi)=-0.8145\sigma^2. \end{aligned}

The covariance changes sign as the lag advances through the cycle, while its oscillations are bounded by the declining envelope $\sigma^2(0.95)^{|h|}$. For instance, that envelope is about half its initial value after $\log(0.5)/\log(0.95)=13.51$ observations. The spectrum and covariance tell a consistent story: variation is concentrated around a cyclical timescale, and dependence persists while oscillating.

The stated frequencies are the centers of the two components, not a claim that the maximum of their sum occurs exactly at either center. In a sum of spectral components, the contribution from one component can shift the maximum of the total density. The interpretation here is about the paired cyclical concentration and its corresponding covariance pattern.

Broad-Band Power and Short Dependence

A broad-band spectrum has substantial variation spread across a wide range of frequencies rather than concentrated in a narrow range. A nearly flat spectrum is the limiting simple case, but broad-band behavior need not be perfectly flat. Smooth variation across frequencies can coexist with autocovariances that are small or vanish after a few lags. The next example illustrates a mild broad-band enhancement rather than a sharp cyclical feature.

Worked Example: A Mildly Frequency-Dependent Economic Series

Suppose a hypothetical stationary series of quarterly price changes is $X_t=\varepsilon_t+0.1\varepsilon_{t-1}$, where the innovations are uncorrelated over time with variance $1$. This is a finite linear filter of the innovations, so its autocovariances are absolutely summable. Its spectral density is

f(\omega)=\frac{1}{2\pi}\left|1+0.1e^{-i\omega}\right|^2 =\frac{1}{2\pi}\left(1.01+0.2\cos\omega\right).

At zero and $\pi$, respectively, the density is

f(0)=\frac{1.21}{2\pi}=0.1926,\qquad f(\pi)=\frac{0.81}{2\pi}=0.1289.

The ratio is $1.21/0.81=1.494$, a modest difference rather than the sharp concentration seen in the persistent-process example. The autocovariances are $\gamma(0)=1.01$, $\gamma(1)=\gamma(-1)=0.1$, and $\gamma(h)=0$ for $|h|>1$. Thus dependence is confined to adjacent observations, and the smooth spectrum varies across the frequency range without a narrow peak. The variance is $1.01$, which also checks against $\int_{-\pi}^{\pi}f(\omega)\,d\omega$.

A Short Justification: Why Positive Dependence Raises the Spectrum at Zero

The zero-frequency result follows directly from the Fourier definition. At $\omega=0$, every exponential factor $e^{-ih\omega}$ equals one, so the spectral density sums the autocovariances without frequency-dependent signs. Because a real stationary process has $\gamma(-h)=\gamma(h)$, the sum separates into lag zero and twice the positive lags:

f(0)=\frac{1}{2\pi}\sum_{h=-\infty}^{\infty}\gamma(h) =\frac{1}{2\pi}\left[\gamma(0)+\sum_{h=1}^{\infty}\{\gamma(h)+\gamma(-h)\}\right] =\frac{1}{2\pi}\left[\gamma(0)+2\sum_{h=1}^{\infty}\gamma(h)\right].

Absolute summability ensures that this Fourier series and the rearrangement are well defined. If many positive-lag covariances are positive, they raise $f(0)$; if they alternate in sign, they may partly cancel. The same principle applies at other frequencies, where the Fourier weights rotate with the lag and can reinforce or offset different covariance patterns.

Reading a Spectrum Carefully

A useful interpretation proceeds from the overall shape, then checks the covariance implications. A high value near zero suggests relatively strong slow-moving variation, but does not by itself establish a trend or nonstationarity. A nonzero-frequency concentration suggests cyclical dependence, but does not prove that the series follows a deterministic cycle. Broad-band power suggests that variance is not dominated by one narrow timescale, but does not mean observations are necessarily independent.

Keep the stationary-process assumption in view. A trend or structural change can produce apparent low-frequency concentration even when a stationary model is inappropriate. Also distinguish the spectrum of a process from a finite-sample estimate: estimates can be noisy, and an apparent peak in a short record may not represent a persistent population feature.

Do not infer a period from a frequency without converting units. The period is $2\pi/\omega$ observations per cycle.
Do not equate a tall, narrow peak with a large total variance. Variance is the area under the spectral density, not its height at a single point.
Do not identify component centers with the exact maximum of a combined spectrum. Other components can shift the maximum.
Do not treat spectral shape as causal evidence. It describes how variation and dependence are distributed across frequencies, not why they arise.

Key takeaway: Low-frequency concentration commonly accompanies sustained positive dependence; cyclical concentration accompanies oscillating autocovariances; and broad-band variation is spread across frequencies rather than dominated by a narrow timescale. Interpret the spectrum alongside the autocovariance and the stationarity assumptions.

Check Your Understanding

Use the frequency–period relation and the connection between spectra and autocovariances to answer the following.

A stationary series has a pronounced concentration near $\omega=0$. What kind of temporal variation does this suggest, and what autocovariance pattern commonly accompanies it?
A spectral feature is centered near $\omega=\pi/5$. What period in observations corresponds to that frequency?
For $\gamma(h)=\sigma^2r^{|h|}\cos(h\omega_0)$, explain what the cosine and the factor $r^{|h|}$ each contribute to the covariance pattern.
Why does a narrow, tall spectral feature not, by itself, establish that the process has high total variance?
Why should one avoid claiming that a component center is necessarily the exact maximum of a sum of spectral components?