You collect 10,000 samples of a stationary noise signal and compute a periodogram. Compared to a 1,000-sample periodogram of the same signal, the 10,000-sample version will have:
AMuch lower variance at each frequency bin, because variance is inversely proportional to sample size just like standard statistical estimators
BFiner frequency resolution due to the denser frequency grid, but the variance at each frequency bin remains approximately the same regardless of N
CBoth lower variance and finer resolution, because more data always improves all aspects of a spectral estimate
DHigher variance, because more samples introduce more noise into the Fourier coefficients
This is the fundamental surprise of spectral estimation. In standard statistics (e.g., estimating a mean), variance decreases as 1/N — more data always means more precision. For the periodogram, each frequency bin corresponds to a single Fourier coefficient computed from all N samples. Increasing N adds new frequencies (finer resolution) but the variance of each individual bin estimate remains approximately S²(k) — the squared true PSD at that frequency. This counterintuitive property is what makes raw periodograms nearly useless for spectral analysis: no matter how much data you collect, the estimate at each frequency bin fluctuates wildly.
Question 2 Multiple Choice
An engineer needs to resolve two spectral peaks separated by 0.5 Hz using a 2-second signal. She considers Welch with 0.5-second segments versus a parametric AR method. What is the key tradeoff?
AWelch with 0.5-second segments has frequency resolution of 1/0.5 = 2 Hz — too coarse to resolve the 0.5 Hz separation; the AR method can achieve super-resolution but only if the AR model and order match the true signal
BThe Welch method is always preferred for this application because parametric methods are computationally too expensive for real-time use
CThe AR method has higher variance than Welch, making it unreliable for resolving closely spaced peaks
DBoth methods provide equivalent frequency resolution; the only tradeoff is computational cost versus implementation simplicity
The frequency resolution of a Welch segment of length M samples is Δf = f_s/M. With 0.5-second segments, Δf = 2 Hz — far too coarse to distinguish two peaks only 0.5 Hz apart. The full 2-second record has 0.5 Hz resolution, but a single periodogram of that record has intolerable variance. An AR parametric method, given correct model order and type, can resolve peaks to arbitrarily fine resolution even from a short record — but if the signal is not well-modeled by an AR process, the estimate may show spurious peaks or miss real features entirely. The choice requires knowing the signal's characteristics.
Question 3 True / False
In the Welch method, using 50% segment overlap produces more independent estimates than 0% overlap, which is why overlap is the recommended strategy for maximum variance reduction.
TTrue
FFalse
Answer: False
Overlapping segments are correlated, not independent. The variance reduction from averaging K overlapping segments is less than K-fold because adjacent segments share data. 0% overlap (fully non-overlapping segments) produces the maximum number of truly independent estimates and the best variance reduction per segment. The reason 50% overlap is commonly recommended is pragmatic: it roughly doubles the number of segments available compared to no overlap, and because the correlation between adjacent segments is moderate, the variance reduction is close to the 2x implied by twice as many segments. It is a practical compromise, not the theoretically optimal strategy for independence.
Question 4 True / False
Parametric spectral estimation methods like AR spectral analysis usually outperform the Welch method when the goal is to identify spectral peaks in a signal.
TTrue
FFalse
Answer: False
Parametric methods offer super-resolution — they can resolve closely spaced peaks that Welch would smear together — but only when the assumed model is correct. If the signal does not follow an AR process, or if the AR order is wrong, the estimated spectrum can show spurious peaks or entirely miss real features. For broadband signals, poorly characterized sources, or when the model order is uncertain, parametric methods can fail catastrophically while Welch remains reliable (if resolution-limited). There is no universally superior method: the choice depends on signal characteristics, available data length, and the consequences of model misspecification.
Question 5 Short Answer
Explain why the periodogram's variance fails to decrease as sample size N increases, and why this is surprising compared to standard statistical estimation.
Think about your answer, then reveal below.
Model answer: In standard estimation (e.g., estimating a mean from N observations), variance decreases as 1/N because each new observation contributes independent information about the same quantity. For the periodogram, the estimate at each frequency bin is based on a single Fourier coefficient — a complex number derived from all N samples. As N increases, the number of frequency bins grows (finer resolution), but the estimate for any particular frequency bin does not benefit from the additional samples in the variance-reduction sense: the bin estimate remains a single noisy draw with variance approximately equal to S²(k). Adding more data creates new bins at finer frequencies rather than averaging down the noise on existing bins. The only way to reduce variance is to average multiple estimates — either across segments (Welch) or by assuming a model (parametric) — not by simply collecting more data.
This fundamental property (inconsistency of the periodogram) was recognized by Bartlett in the 1940s and is what motivates all modern spectral estimation methods. The intuition: a single DFT coefficient is like taking one observation from a distribution — you get information about that frequency, but one observation has the same variance no matter how many samples went into computing it (beyond the minimum needed to define that frequency). Spectral estimation is thus fundamentally different from point estimation: the goal is to control a noise floor that doesn't diminish with N, requiring averaging-based or model-based approaches.