Power spectral density (PSD) is a measurement that shows how the power or energy in a signal is spread across different frequencies. If you have any signal that changes over time, whether it’s an audio recording, a vibration measurement, or brain wave data, PSD breaks it apart and tells you exactly how much power sits at each frequency. Think of it like a prism splitting white light into a rainbow: PSD splits a complex signal into its individual frequency ingredients and shows you how strong each one is.
How PSD Differs From a Simple Power Spectrum
The terms “power spectrum” and “power spectral density” are often used interchangeably, but they measure slightly different things. A power spectrum gives you the power at specific, discrete frequencies. Each value is a standalone number with units like volts squared. A power spectral density, on the other hand, is a density function. Its units are power per unit frequency, such as volts squared per hertz (V²/Hz). The distinction matters: in a PSD, individual values on the curve aren’t the power themselves. Instead, the area under the curve over a frequency range gives you the power in that range. This is similar to how a probability density curve works. The height of the curve at one point isn’t a probability, but the area under a section of the curve is.
This density-based approach makes PSD especially useful for continuous signals or signals measured over long periods, where you care about how power is distributed rather than pinned to a handful of exact frequencies.
What PSD Actually Tells You
A PSD plot has frequency on the horizontal axis and power density on the vertical axis. Tall peaks indicate frequencies where the signal carries a lot of energy. A flat region means energy is spread evenly across those frequencies. The noise floor, the low baseline visible across the plot, represents the background level of power where no dominant signal exists.
If you add a pure tone (a single-frequency sound) to a recording, it shows up as a sharp spike at that frequency. Add a second tone and you get a second spike. Real-world signals are messier: they produce broader peaks, harmonic patterns (evenly spaced spikes at multiples of a base frequency), and complex shapes. Electrical interference from power lines, for instance, shows up as sharp spikes at 50 or 60 Hz and at their harmonics (100, 150, 200 Hz, and so on).
The carrier frequency is the frequency with the highest power level in the spectrum, and it often represents the dominant behavior in whatever system you’re measuring.
Getting Total Power From a PSD
One of the most practical features of PSD is that you can recover the total power in a signal by integrating (summing up the area under) the entire PSD curve. If you only care about a specific frequency band, you integrate over just that range. This is a direct consequence of Parseval’s theorem, which guarantees that the total energy you’d calculate by squaring the signal in the time domain equals the total energy you get by summing up contributions across all frequencies in the frequency domain. In equation form, the average power of a signal equals the integral of its PSD over all frequencies.
This is not just a mathematical convenience. It’s how engineers measure the power of noise in a communication channel, how physicists quantify vibration energy in a structure, and how clinicians compare brain activity across frequency bands.
How PSD Is Calculated
You can’t measure PSD directly. You estimate it from a finite chunk of recorded data. The simplest approach is the periodogram: take a block of data, compute its Fourier transform (which converts the signal from time to frequency), square the result, and normalize by the recording length and sampling rate. The formula boils down to dividing the squared magnitude of the Fourier transform by the product of the number of samples and the sampling frequency.
The problem with a single periodogram is that it’s noisy. Small random fluctuations in your data produce jagged, unreliable estimates. Welch’s method solves this by splitting the signal into overlapping blocks, computing a periodogram for each block, and then averaging them all together. The averaging smooths out the randomness, giving you a much more stable estimate. When windowing functions other than a simple rectangular cutoff are applied to each block, the blocks typically overlap (often by 50%) so that no data near the edges is lost.
Another common approach uses autoregressive models, which fit a mathematical model to the signal’s autocorrelation structure and then derive the PSD from that model. This method, sometimes called the Burg method, produces smoother spectra and works well for short data segments.
Underneath all of these methods sits the Wiener-Khinchin theorem, which states that a signal’s PSD is the Fourier transform of its autocovariance function. The autocovariance measures how similar a signal is to a time-shifted version of itself. Frequencies where the signal is highly self-similar show up as strong PSD peaks. This theorem provides the theoretical foundation that connects the statistical properties of a signal to its frequency content.
Units You’ll Encounter
PSD units always take the form of “signal units squared per hertz.” For a voltage signal, that’s V²/Hz. For an acceleration signal from a vibration sensor, it might be (m/s²)²/Hz, often written as g²/Hz. In geophysics, you might see units like watts per kilogram per hertz (W/kg/Hz).
Because the numbers can span many orders of magnitude, PSD is frequently plotted on a logarithmic scale in decibels. The conversion is 10 times the base-10 logarithm of the power density, yielding units like dB/Hz. For EEG brain wave data, for example, PSD output is commonly expressed as 10·log₁₀(μV²/Hz). Logarithmic scaling makes it much easier to see low-power details that would be invisible on a linear plot.
Noise Types and PSD Slopes
PSD plots reveal the character of noise at a glance. When plotted on a log-log scale (both axes logarithmic), different noise types produce straight lines with characteristic slopes, all following a general 1/f^α pattern where α determines the slope.
- White noise has a flat PSD (α = 0), meaning equal power at every frequency. Radio static is a close real-world example.
- Pink noise has a PSD that drops at 3 dB per octave (α = 1), so lower frequencies carry more power. It sounds like a waterfall or steady rain and is common in electronic circuits and natural systems.
- Brown (Brownian) noise falls off more steeply at 6 dB per octave (α = 2), with even more energy concentrated at low frequencies. It sounds like deep rumbling.
Identifying the slope of a PSD plot is a quick way to classify the type of process generating a signal, which has implications in fields from climate science to financial modeling.
PSD in Brain Wave Analysis
One of the most developed clinical applications of PSD is in electroencephalography (EEG), where it quantifies the electrical activity of cortical brain cells across standard frequency bands: delta (0.5–4 Hz), theta (4–7 Hz), alpha1 (8–10 Hz), alpha2 (10–12 Hz), beta (13–30 Hz), and gamma (30–40 Hz). By computing the relative PSD in each band, clinicians can see how brain activity is distributed and detect abnormal patterns.
In Alzheimer’s disease research, PSD analysis has been particularly revealing. Patients with Alzheimer’s show a characteristic shift: power increases in the slower theta and delta bands while decreasing in the faster alpha and beta bands. This “slowing” of the EEG signal correlates with disease severity. Studies have found statistically significant differences in the theta and alpha2 bands between Alzheimer’s patients and healthy controls, especially over the parietal, temporal, and occipital regions of the scalp. The amount of power redistribution toward lower frequencies tracks with how far the disease has progressed, making PSD a quantitative tool for monitoring neurological decline rather than relying solely on subjective clinical assessments.
Other Common Applications
PSD shows up wherever signals vary over time and people need to understand their frequency content. In telecommunications, engineers use PSD to measure how a transmitted signal’s power is distributed across a bandwidth, ensuring it stays within regulatory limits and doesn’t interfere with adjacent channels. In mechanical engineering, vibration PSD profiles are used to design products that can withstand the shaking they’ll experience during transport or operation. Seismologists analyze the PSD of ground motion recordings to characterize earthquake energy. Audio engineers use PSD to analyze the spectral content of recordings and identify unwanted resonances or interference.
In each case, the core idea is the same: break a complex, time-varying signal into its frequency components, quantify how much power each frequency carries, and use that information to make decisions about the system producing the signal.