Ferkans — Interactive Telecom Tutor

Why Windows Matter

When you compute the FFT of a finite signal, you implicitly multiply it by a rectangular window. This abrupt truncation causes spectral leakage: energy from a single frequency spreads across many bins. Window functions taper the signal's edges to zero, trading main-lobe width (resolution) for sidelobe suppression (dynamic range). Choosing the right window is the key design decision in spectral analysis.

Definition:
Window Functions

A window function $w[n]$ is a finite-length sequence that is multiplied with a signal before computing the DFT to reduce spectral leakage. The windowed DFT is:

$X_w[k] = \sum_{n=0}^{N-1} w[n]\, x[n]\, e^{-j2\pi kn/N}$

Common windows in scipy.signal.windows:

Window	Main-Lobe Width	Sidelobe Level	Use Case
Rectangular	Narrowest ( $2/N$ )	$-13$ dB	Maximum resolution
Hann	$4/N$	$-31$ dB	General-purpose
Hamming	$4/N$	$-43$ dB	Speech/audio
Blackman	$6/N$	$-58$ dB	High dynamic range
Kaiser ( $\beta$ )	Adjustable	Adjustable	Flexible trade-off

from scipy.signal import windows
w = windows.hann(N)
X = np.fft.fft(w * x)

Definition:
The Kaiser Window Family

The Kaiser window provides a continuous trade-off between main-lobe width and sidelobe attenuation via the parameter $\beta$ :

$w[n] = \frac{I_0\!\left(\beta\sqrt{1 - \left(\frac{2n}{N-1} - 1\right)^2}\right)}{I_0(\beta)}$

where $I_0$ is the zeroth-order modified Bessel function.

$\beta$	Sidelobe Attenuation	Equivalent Window
0	$-13$ dB	Rectangular
5	$-36$ dB	Hamming-like
8.6	$-60$ dB	Blackman-like

w = windows.kaiser(N, beta=8.6)

Definition:
Periodogram Power Spectral Density Estimate

The periodogram estimates the power spectral density (PSD) as:

$\hat{S}_{xx}(f_k) = \frac{1}{N f_s} |X[k]|^2$

In SciPy:

from scipy.signal import periodogram
f, Pxx = periodogram(x, fs=1000, window='hann')

The periodogram is the simplest PSD estimator but has high variance: its standard deviation equals its mean, regardless of signal length.

The periodogram is an inconsistent estimator — increasing $N$ does not reduce its variance. Use Welch's method for a smoother estimate.

Definition:
Welch's Method for PSD Estimation

Welch's method reduces the variance of the periodogram by:

Dividing the signal into overlapping segments
Windowing each segment
Computing the periodogram of each segment
Averaging the periodograms

from scipy.signal import welch
f, Pxx = welch(x, fs=1000, nperseg=256, noverlap=128, window='hann')

With $K$ segments and 50% overlap, the variance is reduced by approximately a factor of $9K/11$ . The trade-off: averaging reduces variance but also reduces frequency resolution to $\Delta f = f_s / L$ where $L$ is the segment length.

Definition:
Short-Time Fourier Transform (STFT)

The STFT applies the DFT to successive windowed segments of a signal, producing a time-frequency representation:

$\text{STFT}\{x\}(m, k) = \sum_{n=0}^{L-1} w[n]\, x[n + mH]\, e^{-j2\pi kn/L}$

where $L$ is the window length and $H$ is the hop size.

from scipy.signal import stft, spectrogram
f, t, Zxx = stft(x, fs=1000, nperseg=256, noverlap=192)
f, t, Sxx = spectrogram(x, fs=1000, nperseg=256, noverlap=192)

spectrogram returns $|Zxx|^2$ , the squared magnitude of the STFT.

The STFT has a fixed time-frequency resolution determined by the window length: long windows give good frequency resolution but poor time resolution, and vice versa. This is a fundamental limitation captured by the uncertainty principle.

Theorem: Time-Frequency Uncertainty Principle

No signal can be simultaneously localized in both time and frequency. For any signal $x(t)$ , the time spread $\sigma_t$ and frequency spread $\sigma_f$ satisfy:

$\sigma_t \cdot \sigma_f \ge \frac{1}{4\pi}$

Equality holds only for the Gaussian pulse $x(t) = e^{-\alpha t^2}$ .

A short pulse (good time localization) must contain many frequencies (poor frequency localization). A pure sinusoid (perfect frequency localization) extends infinitely in time. The STFT window length controls where you sit on this trade-off.

Example: Effect of Window Choice on Spectral Analysis

A signal contains a strong tone at 100 Hz and a weak tone at 105 Hz that is 40 dB below. Compare the ability of different windows (rectangular, Hann, Blackman) to reveal the weak tone.

Solution

Generate signal and apply windows

from scipy.signal import windows
import numpy as np

fs = 1000
t = np.arange(0, 1, 1/fs)
x = np.sin(2*np.pi*100*t) + 0.01*np.sin(2*np.pi*105*t)

for name, w_func in [("Rectangular", windows.boxcar),
                      ("Hann", windows.hann),
                      ("Blackman", windows.blackman)]:
    w = w_func(len(x))
    X = np.fft.fft(w * x, n=8192)
    mag_db = 20 * np.log10(np.abs(X[:4096]) / np.max(np.abs(X)))

The Blackman window's $-58$ dB sidelobes reveal the weak 105 Hz tone, while the rectangular window's $-13$ dB sidelobes mask it.

Example: Welch's Method vs Periodogram

Compare the periodogram and Welch's method for estimating the PSD of a signal consisting of two sinusoids in white noise.

Solution

Compute both estimates

from scipy.signal import periodogram, welch
import numpy as np

fs = 1000
t = np.arange(0, 10, 1/fs)
x = np.sin(2*np.pi*50*t) + np.sin(2*np.pi*120*t) + np.random.randn(len(t))

f_per, P_per = periodogram(x, fs=fs, window='hann')
f_welch, P_welch = welch(x, fs=fs, nperseg=1024, noverlap=512)

The Welch estimate is much smoother, clearly showing the two spectral peaks above the noise floor, while the periodogram is noisy and harder to interpret.

Example: Spectrogram of a Linear Chirp

Generate a linear chirp signal from 100 Hz to 400 Hz over 2 seconds and compute its spectrogram to visualize the time-frequency content.

Solution

Generate and analyze

from scipy.signal import chirp, spectrogram
import numpy as np

fs = 2000
t = np.arange(0, 2, 1/fs)
x = chirp(t, f0=100, f1=400, t1=2, method='linear')

f, t_spec, Sxx = spectrogram(x, fs=fs, nperseg=256, noverlap=240)

The spectrogram shows a bright diagonal line from (0 s, 100 Hz) to (2 s, 400 Hz), clearly revealing the linear frequency sweep.

Window Function Comparison

Compare different window functions in both the time domain and frequency domain. Observe the trade-off between main-lobe width and sidelobe level.

Parameters

Spectral Analysis Recipes

python

Complete examples: window functions, periodogram, Welch PSD, spectrogram, and STFT with various windowing options.

# Code from: ch07/python/spectral_analysis.py
# Load from backend supplements endpoint

Quick Check

Welch's method reduces the variance of the PSD estimate by:

Using a longer FFT

Averaging periodograms of overlapping windowed segments

Applying zero-padding

Using a wider window

Correction:

Averaging periodograms of overlapping windowed segments

Averaging $K$ independent (or partially overlapping) periodograms reduces variance by approximately $K$ .

Common Mistake: Using Rectangular Window for Spectral Analysis

Mistake:

Computing the FFT without applying a window function first. This implicitly uses a rectangular window, which has $-13$ dB sidelobes that can mask weak spectral components near strong ones.

Correction:

Always apply a window before the FFT for spectral analysis:

from scipy.signal import windows
w = windows.hann(len(x))
X = np.fft.fft(w * x)

Key Takeaway

Window functions trade frequency resolution for sidelobe suppression. Use Welch's method instead of the periodogram for smoother PSD estimates. The spectrogram (STFT magnitude squared) reveals time-varying frequency content but has fixed time-frequency resolution governed by the uncertainty principle.

Power Spectral Density (PSD)

The distribution of signal power as a function of frequency, measured in units of power per Hz (e.g., V $^2$ /Hz).

STFT (Short-Time Fourier Transform)

A time-frequency representation obtained by applying the DFT to successive windowed segments of a signal.

Related: Spectrogram, window function

Why This Matters: Spectrum Sensing in Cognitive Radio

Cognitive radios use spectral analysis to detect whether a frequency band is occupied by a primary user. The Welch periodogram is a standard approach: if the estimated PSD in a band exceeds a threshold, the band is declared occupied. The trade-off between detection probability and false alarm rate maps directly to the segment length and overlap parameters. Window choice affects the ability to detect weak signals near strong interferers.

See full treatment in Chapter 22

Window Functions at a Glance

Window	Main-Lobe Width	Peak Sidelobe	Best For
Rectangular	$2\Delta f$	$-13$ dB	Maximum resolution, no dynamic range needed
Hann	$4\Delta f$	$-31$ dB	General-purpose spectral analysis
Hamming	$4\Delta f$	$-43$ dB	Speech processing, moderate dynamic range
Blackman	$6\Delta f$	$-58$ dB	High dynamic range, weak signal detection
Kaiser ( $\beta=8.6$ )	$\sim 5\Delta f$	$-60$ dB	Flexible trade-off via $\beta$

Window Functions and Spectral Analysis

Why Windows Matter

Definition: Window Functions

Definition: The Kaiser Window Family

Definition: Periodogram Power Spectral Density Estimate

Definition: Welch's Method for PSD Estimation

Definition: Short-Time Fourier Transform (STFT)

Theorem: Time-Frequency Uncertainty Principle

Example: Effect of Window Choice on Spectral Analysis

Generate signal and apply windows

Example: Welch's Method vs Periodogram

Compute both estimates

Example: Spectrogram of a Linear Chirp

Generate and analyze

Window Function Comparison

Parameters

Spectral Analysis Recipes

Quick Check

Common Mistake: Using Rectangular Window for Spectral Analysis

Key Takeaway

Power Spectral Density (PSD)

STFT (Short-Time Fourier Transform)

Why This Matters: Spectrum Sensing in Cognitive Radio

Window Functions at a Glance

Definition:
Window Functions

Definition:
The Kaiser Window Family

Definition:
Periodogram Power Spectral Density Estimate

Definition:
Welch's Method for PSD Estimation

Definition:
Short-Time Fourier Transform (STFT)