Questions: Digital Spectral Analysis: Nonparametric Methods

Question 1

You have 10,000 samples of a signal. You compute a periodogram. Then you collect 100,000 samples of the same signal and compute another periodogram. How does the statistical variance of the spectral estimate change?

• A. It decreases by a factor of 10, since you have 10× more data
• B. It decreases slightly due to improved frequency resolution
• C. It does not decrease — the periodogram is an inconsistent estimator whose variance is independent of record length
• D. It doubles, because more data introduces more spectral leakage

Answer: (C)

The periodogram is an inconsistent estimator: each DFT bin has approximately a chi-squared distribution with 2 degrees of freedom regardless of how many samples you collect. More data gives finer frequency resolution (smaller bin spacing), but the variance in each bin does not decrease. You get more bins, not less noisy bins. This is the fundamental weakness of the periodogram and the motivation for Welch and multitaper methods, which achieve variance reduction through averaging, not longer records.

Question 2

In the Welch method, you shorten the segment length M while keeping the total record length N fixed (so more segments are averaged). What is the tradeoff?

• A. Variance decreases and frequency resolution improves — shortening segments is always better
• B. Variance decreases because more segments are averaged, but frequency resolution degrades because each segment is shorter
• C. Variance stays the same because the total data is unchanged, but resolution improves
• D. Both variance and resolution worsen because shorter segments cause more spectral leakage

Answer: (B)

The Welch method's central design tension: frequency resolution is Δf = f_s/M (determined by segment length M), while variance is reduced by averaging L ≈ N/M segments. Shorter M → more segments → lower variance, but also coarser resolution (can't distinguish frequencies closer than f_s/M). Longer M → finer resolution, but fewer segments to average → higher variance. The engineer must choose M to balance these competing demands based on which spectral features matter most for the application.

Question 3

The periodogram is a consistent spectral estimator: as the number of data samples N increases, its variance at each frequency bin decreases toward zero.

• T. True
• F. False

Answer: False

This is the critical misconception about the periodogram. It is provably inconsistent — its variance does not decrease with N. Each DFT bin of the periodogram has an approximately chi-squared(2) distribution, giving a coefficient of variation near 1 (standard deviation ≈ mean) regardless of record length. Longer records produce more frequency bins (finer resolution), but each individual bin estimate remains just as noisy. Consistency requires either averaging (Welch) or multiple estimates (multitaper); the raw periodogram provides neither.

Question 4

The multitaper method achieves variance reduction compared to the periodogram by applying multiple orthogonal Slepian tapers to the full data record and averaging the resulting spectral estimates.

• T. True
• F. False

Answer: True

The multitaper method's insight is that you can extract K nearly independent spectral estimates from the same N-sample record by using K orthogonal Slepian tapers, each optimized to concentrate energy within a specified bandwidth W. Averaging K independent estimates reduces variance by approximately K — similar to Welch's averaging, but without shortening the record (and thus without sacrificing full-record frequency resolution for the averaged estimate). The price is that spectral features closer than W Hz are blurred. This is why multitaper is preferred for high dynamic-range signals where leakage from dominant peaks would contaminate nearby weaker features.

Question 5

Explain the fundamental tension between spectral resolution and variance in nonparametric spectral estimation, and why simply collecting more data does not resolve it.

Model answer: Resolution is determined by record or segment length: finer resolution requires longer segments (Welch) or larger bandwidth-time products (multitaper). Variance reduction requires averaging independent spectral estimates. These compete because averaging requires either shortening segments (reducing resolution) or using multiple tapers (blurring features within bandwidth W). More data helps only if used strategically — by enabling longer segments with more averaging, or by justifying a larger NW product. A longer raw periodogram gives finer resolution but the same per-bin variance, because each bin is still one chi-squared(2) estimate.

This tension is intrinsic to the spectral estimation problem: each DFT bin is formed from sinusoidal basis functions with a given frequency resolution, and a single estimate from that basis function is inherently noisy. To reduce noise, you need multiple independent looks at the spectrum — but getting independent looks from a single stationary signal requires either dividing the record (sacrificing resolution) or using orthogonal tapers (accepting blurring). There is no free lunch: you cannot simultaneously have high resolution and low variance from a finite record.