14.3 Spectral Estimation Techniques

Spectral estimation techniques let you figure out how a signal's power is spread across different frequencies. This is the practical side of Fourier analysis: taking real, finite data and extracting meaningful frequency-domain information from it. You'll encounter two broad categories of methods here (non-parametric and parametric), and understanding when to use each one is just as important as knowing how they work.

Power Spectral Density Estimation

Fourier-based Methods for PSD Estimation

The power spectral density (PSD) describes how a signal's power is distributed over frequency. Think of it as a map showing which frequencies carry the most energy. Fourier-based methods estimate the PSD directly from data without assuming any particular signal model.

The simplest approach is the periodogram. You take the Discrete Fourier Transform (DFT) of your signal, compute the squared magnitude at each frequency bin, and normalize by the signal length $N$:

$\hat{S}(f) = \frac{1}{N} |X(f)|^2$

This gives you a raw estimate of the PSD. The problem? Periodograms are noisy. The variance of the estimate doesn't decrease as you collect more data, which makes the result look jagged and unreliable.

Welch's method fixes this by breaking the signal into overlapping segments, computing a periodogram for each one, and averaging them together. The steps are:

1. Divide the signal into segments of length $L$, with a specified overlap (commonly 50%)
2. Apply a window function (e.g., Hann or Hamming) to each segment
3. Compute the periodogram of each windowed segment
4. Average all the segment periodograms to produce the final PSD estimate

The averaging is what drives the variance down. More segments means more averaging and a smoother estimate.

Factors Influencing Spectral Estimate Quality

Several choices affect how good your spectral estimate turns out:

• Signal length and zero-padding: Longer signals give you finer frequency resolution because the spacing between DFT bins is $\Delta f = \frac{f_s}{N}$, where $f_s$ is the sampling rate. Zero-padding interpolates between bins (making the spectrum look smoother) but doesn't actually add new information or improve true resolution.
• Window functions: Applying a window like Hann or Hamming before computing the DFT reduces spectral leakage, which is the smearing of energy from one frequency into neighboring bins. Different windows trade off between main-lobe width (resolution) and side-lobe level (leakage suppression).
• Segment length in Welch's method: Longer segments give better frequency resolution but fewer segments to average, so variance stays higher. Shorter segments give more averaging (lower variance) but coarser frequency resolution. This is a fundamental trade-off.
• Overlap: Overlapping segments (typically 50%) lets you squeeze more segments out of the same data, providing additional averaging without needing a longer recording.

Parametric vs Non-parametric Spectral Estimation

Non-parametric Spectral Estimation Methods

Non-parametric methods estimate the PSD directly from the data without assuming any underlying signal model. The periodogram and Welch's method both fall in this category.

Advantages:

• Computationally straightforward (just FFTs and averaging)
• No prior knowledge of the signal's structure is needed
• Work reasonably well for a wide range of signal types

Limitations:

• Can suffer from high variance (periodogram) or reduced resolution (Welch's method with short segments)
• Sensitive to the choice of window function and segment parameters
• Need relatively long data records to achieve good frequency resolution

Parametric Spectral Estimation Methods

Parametric methods assume the signal was generated by a specific model, typically an autoregressive (AR) model, and estimate the PSD by fitting that model's parameters to the data. The two most common approaches are:

• Yule-Walker method: Estimates AR model coefficients by solving a system of equations based on the signal's autocorrelation values.
• Burg method: Estimates AR coefficients by minimizing the forward and backward prediction error simultaneously, which tends to produce more stable results for short data records.

Once you have the AR model parameters, you compute the PSD analytically from the model rather than from the raw DFT.

Advantages:

• Can achieve much higher frequency resolution than non-parametric methods, especially with short data records
• Produce smoother estimates with lower variance
• Particularly effective for signals with sharp, well-defined spectral peaks

Limitations:

• If the assumed model doesn't match the actual signal, the estimate can be misleading
• Sensitive to model order selection (how many AR coefficients to use). Too few coefficients over-smooth the spectrum; too many introduce spurious peaks.
• Struggle with complex or non-stationary signals that don't fit a simple AR structure

Spectral Estimation Methods Evaluation

Resolution

Resolution is the ability to distinguish two closely spaced frequency components. If two sinusoids are close in frequency, a low-resolution method will merge them into a single broad peak, while a high-resolution method will show two distinct peaks.

The periodogram's resolution is set by the data length: $\Delta f \approx \frac{1}{T}$, where $T$ is the total observation time. For short records, this can be quite coarse. Welch's method trades some of that resolution away (because each segment is shorter than the full record) in exchange for lower variance.

Parametric methods can resolve closely spaced frequencies even with short data, which is one of their biggest practical advantages.

Bias and Variance

• Bias is the systematic error between your estimate's expected value and the true PSD. Window functions and segment averaging can both introduce bias. An estimator is called unbiased if its expected value equals the true PSD as the number of samples grows.
• Variance is how much the estimate fluctuates from one realization of the signal to another. Lower variance means more consistent, trustworthy results.

Here's how the methods compare:

The periodogram is a high-variance estimator. Welch's method brings variance down through averaging. Parametric methods can achieve low variance, but only when the model is a good fit for the data.

Trade-offs and Considerations

Every spectral estimation choice involves trade-offs:

• Smoothing vs. resolution: Bartlett's method (non-overlapping averaged periodograms) and Welch's method both reduce variance by averaging, but at the cost of frequency resolution. You can't have both with finite data.
• Welch's parameters: Window type, segment length, and overlap percentage all interact. There's no single "best" setting; it depends on whether you care more about resolving close frequencies or getting a stable estimate.
• Parametric model order: This is the critical decision for AR-based methods. Formal criteria can help:
  • Akaike Information Criterion (AIC): Balances model fit against complexity, penalizing extra parameters
  • Bayesian Information Criterion (BIC): Similar to AIC but applies a stronger penalty for additional parameters, favoring simpler models You typically compute AIC or BIC for a range of model orders and pick the order that minimizes the criterion.

Spectral Estimate Interpretation

Identifying Dominant Frequencies and Signal Characteristics

Once you have a PSD estimate, reading it is about spotting patterns:

• Peaks correspond to dominant periodic components. A sharp peak at 50 Hz in an electrical signal's PSD, for example, points to power line interference at that frequency.
• Peak width tells you about the stability of a component. A narrow peak means a well-defined, stable oscillation. A broad peak suggests the frequency is drifting over time or is mixed with noise.
• Peak height reflects relative power. Taller peaks carry more energy than shorter ones, so you can rank frequency components by their importance to the overall signal.

Analyzing Signal Characteristics and Patterns

The PSD's overall shape reveals the signal's nature beyond just individual peaks:

• White noise has a flat PSD, meaning equal power at all frequencies.
• 1/f noise (pink noise) has a PSD that decreases with frequency, appearing as a downward slope on a log-log plot. This pattern shows up in many natural and electronic systems.
• Multiple peaks can indicate several distinct oscillatory processes happening at once. In an EEG brain signal, for instance, separate peaks might correspond to alpha rhythms (~8-13 Hz), beta rhythms (~13-30 Hz), and gamma rhythms (~30-100 Hz), each associated with different brain states.

Comparing PSD estimates across different conditions is one of the most practical applications. If you record vibration data from a machine under normal operation and then under a fault condition, new peaks or shifts in existing peaks can pinpoint the fault's frequency signature. The same logic applies to signal classification, anomaly detection, and studying how a system responds to changes.