Questions: Parametric Signal Models: AR, MA, and ARMA

Question 1

An engineer uses an AR(2) model to represent a vibration signal with a single narrow resonance peak. Compared to computing the 64-point DFT of the same data, what is the key advantage of the parametric approach?

• A. The AR model is computationally cheaper because it requires fewer multiplications than an FFT
• B. The AR model can resolve narrower spectral peaks from short data records because it encodes structure as filter parameters rather than raw periodogram bins
• C. The AR model is more robust to noise because it smooths the spectrum automatically
• D. The AR model applies to non-stationary signals whereas the DFT assumes stationarity

Answer: (B)

The fundamental advantage of parametric methods is frequency resolution per data point. A 64-point DFT has frequency resolution of f_s/64 — two peaks closer than this appear as one blurred lump. An AR(2) model captures a resonance with just 2 parameters and can resolve peaks far sharper than any DFT of the same data length, because it encodes prior structure (the signal comes from a pole-zero system driven by white noise). This is the key tradeoff: if the model is correct, parametric methods massively outperform nonparametric ones. If the model is wrong, the result is misleading.

Question 2

A researcher fits AR models of increasing order to a short data segment. At order p=15 the prediction error variance is still decreasing. What is the primary risk of continuing to increase the order to p=30?

• A. The Levinson-Durbin algorithm becomes numerically unstable at high orders, producing complex-valued coefficients
• B. The model overfits: it begins fitting the noise structure as if it were signal, generating spurious spectral peaks at meaningless frequencies
• C. The model underfits because AR models of order greater than 20 cannot represent spectral peaks below 1 kHz
• D. Higher orders require more data to estimate, but the resolution improves proportionally

Answer: (B)

Overfitting is the central risk of high model order. At some point, additional poles no longer represent true resonances in the signal — they fit the particular noise realization in the data record. The resulting spectral estimate shows sharp peaks that are artifacts of the finite sample, not properties of the underlying process. The AIC and MDL criteria formalize this tradeoff by penalizing complexity: the optimal order is where reduced fitting error no longer justifies the added parameters. Spurious peaks from overfit AR models have led to false physical interpretations in seismology, biomedical signal analysis, and other fields.

Question 3

An AR model is better suited than an MA model for representing a signal with a sharp spectral resonance, because all-poles models can efficiently capture narrow peaks with few parameters.

• T. True
• F. False

Answer: True

An AR(p) model is an all-poles system — its transfer function H(z) = 1/A(z) has poles but no zeros. Poles create resonances in the spectral estimate. A single pair of complex conjugate poles close to the unit circle creates a sharp narrow peak in the power spectrum. Representing the same narrow peak with an MA model (all zeros) requires very high order because zeros produce spectral nulls, not peaks; you need many zeros arranged to create a peak by cancellation. For signals dominated by resonances (speech formants, EEG rhythms, vibrating structures), AR models are the natural, parsimonious choice.

Question 4

AR, MA, and ARMA models are different types of signals — AR signals have different fundamental properties than MA signals and cannot be represented by each other.

• T. True
• F. False

Answer: False

AR, MA, and ARMA are MODELS, not signal types. An underlying physical process doesn't 'know' which model is being applied to it. In fact, any ARMA process can be approximated arbitrarily well by a sufficiently high-order AR model (due to the all-poles approximation of rational spectra). The choice of model family is about parsimony: for a signal dominated by resonances, AR captures the structure efficiently; for signals with spectral notches, MA is efficient; for signals with both, ARMA is most compact. This is a modeling choice, not a description of the signal's intrinsic nature.

Question 5

Why can a parametric AR model resolve two closely spaced frequency components that would appear as a single blurred peak in a nonparametric periodogram of the same data length?

Model answer: A periodogram's frequency resolution is fundamentally limited by data length (approximately f_s/N for N samples). Two sinusoids within this resolution bin appear merged. An AR model bypasses this limit by assuming the signal is generated by a system with specific poles — it encodes the spectral structure as model parameters rather than DFT bins. Two closely spaced resonances are captured by two pole pairs close to the unit circle, regardless of data length. This extrapolates beyond the Fourier limit by exploiting the structural assumption. The risk is that if the model assumption is wrong, spurious resolution appears as sharp artifactual peaks.

This resolution advantage is why AR spectral analysis is used in applications where data is inherently short: radar Doppler processing (few range gate samples), neural signal analysis (short stationary epochs), and rotating machinery monitoring (need resolution before failure progresses). The price paid is that the approach is only valid when the data genuinely comes from a pole-zero process driven by white noise — a parametric assumption with no free lunch.