Login
OR
Don't worry. We'll fix this for you!
In electronics and telecommunication engineering, signals often contain random or unpredictable components due to noise, interference, or inherent randomness in the source. Such signals are modeled as random processes, which are collections of random variables indexed by time. Understanding and analyzing these random signals is crucial for designing reliable communication systems, noise reduction techniques, and signal detection algorithms.
Two fundamental tools for analyzing random processes are autocorrelation and power spectral density (PSD). Autocorrelation provides insight into how a signal relates to itself over time shifts, revealing patterns and memory within the signal. PSD, on the other hand, describes how the signal's power is distributed across different frequencies, helping engineers understand the frequency content of random signals.
This section will build your understanding of these concepts from the ground up, starting with the nature of random processes, moving through the mathematical definitions and properties of autocorrelation, and culminating in the interpretation and calculation of power spectral density. Along the way, practical examples and diagrams will clarify these ideas, preparing you for both theoretical and exam challenges.
Imagine you are listening to a piece of music. If you play the same segment again after a short pause, you will notice how similar or different it sounds compared to the original. This idea of comparing a signal with a time-shifted version of itself is the essence of autocorrelation.
Definition: Autocorrelation measures the similarity between a signal and a delayed (or shifted) copy of itself as a function of the delay, called the lag. For a random process \( X(t) \), the autocorrelation function (ACF) is defined as the expected value of the product of the signal at time \( t \) and at time \( t + \tau \):
Here, \( E[\cdot] \) denotes the statistical expectation, which averages over all possible realizations of the random process. The lag \( \tau \) can be positive or negative, representing shifts forward or backward in time.
Properties of Autocorrelation Function:
For discrete-time random signals \( X_n \), the autocorrelation at lag \( k \) is:
To visualize autocorrelation, consider the following diagram:
In this illustration, the blue curve represents the original signal \( X(t) \), the dashed blue curve is the time-shifted signal \( X(t + \tau) \), and the red curve shows the autocorrelation function \( R_X(\tau) \) plotted against lag \( \tau \). Notice how the autocorrelation peaks at zero lag and decreases as the lag increases, indicating less similarity as the signals shift further apart.
While autocorrelation describes how a signal correlates with itself over time, engineers often need to understand how the signal's power is distributed across different frequencies. This is where the power spectral density (PSD) comes in.
Definition: The PSD of a random process \( X(t) \), denoted \( S_X(f) \), represents the distribution of power per unit frequency. It tells us how much power lies within each frequency band.
The fundamental link between autocorrelation and PSD is given by the Wiener-Khinchin theorem, which states that the PSD is the Fourier transform of the autocorrelation function:
Conversely, the autocorrelation function can be recovered by the inverse Fourier transform of the PSD:
This duality means that time-domain correlation information and frequency-domain power distribution are two sides of the same coin.
To visualize PSD, consider the following plot of a random process's power distribution over frequency:
In this plot, the green curve represents the PSD \( S_X(f) \), showing how power is concentrated around certain frequencies. Peaks indicate dominant frequency components, while flat regions indicate uniform power distribution.
Step 1: Understand the definition of discrete autocorrelation:
\( R_X[k] = E[X_n X_{n+k}] \). For a finite sequence, we approximate expectation by averaging over available products.
Step 2: Calculate \( R_X[0] \) (zero lag):
\( R_X[0] = \frac{1}{4} (2^2 + (-1)^2 + 3^2 + 0^2) = \frac{1}{4} (4 + 1 + 9 + 0) = \frac{14}{4} = 3.5 \)
Step 3: Calculate \( R_X[1] \) (lag 1):
Multiply pairs separated by 1:
\( (2)(-1) + (-1)(3) + (3)(0) = -2 -3 + 0 = -5 \)
Number of pairs = 3, so average:
\( R_X[1] = \frac{-5}{3} \approx -1.67 \)
Step 4: Calculate \( R_X[2] \) (lag 2):
Multiply pairs separated by 2:
\( (2)(3) + (-1)(0) = 6 + 0 = 6 \)
Number of pairs = 2, so average:
\( R_X[2] = \frac{6}{2} = 3 \)
Answer: The autocorrelation values are \( R_X[0] = 3.5 \), \( R_X[1] = -1.67 \), and \( R_X[2] = 3 \).
Step 1: Recall the PSD is the Fourier transform of \( R_X(\tau) \):
\( S_X(f) = \int_{-\infty}^{\infty} e^{-\alpha |\tau|} e^{-j 2 \pi f \tau} d\tau \)
Step 2: Split the integral into positive and negative parts:
\( S_X(f) = \int_{0}^{\infty} e^{-\alpha \tau} e^{-j 2 \pi f \tau} d\tau + \int_{-\infty}^{0} e^{\alpha \tau} e^{-j 2 \pi f \tau} d\tau \)
Step 3: Evaluate each integral:
For \( \tau \geq 0 \):
\( I_1 = \int_0^\infty e^{-\alpha \tau} e^{-j 2 \pi f \tau} d\tau = \int_0^\infty e^{-(\alpha + j 2 \pi f) \tau} d\tau = \frac{1}{\alpha + j 2 \pi f} \)
For \( \tau < 0 \):
Let \( u = -\tau \), then \( \tau = -u \), \( d\tau = -du \), limits change from \( \tau: -\infty \to 0 \) to \( u: \infty \to 0 \):
\( I_2 = \int_{-\infty}^0 e^{\alpha \tau} e^{-j 2 \pi f \tau} d\tau = \int_\infty^0 e^{-\alpha u} e^{j 2 \pi f u} (-du) = \int_0^\infty e^{-\alpha u} e^{j 2 \pi f u} du = \frac{1}{\alpha - j 2 \pi f} \)
Step 4: Sum the two integrals:
\( S_X(f) = \frac{1}{\alpha + j 2 \pi f} + \frac{1}{\alpha - j 2 \pi f} = \frac{2 \alpha}{\alpha^2 + (2 \pi f)^2} \)
Answer: The PSD is
\[ S_X(f) = \frac{2 \alpha}{\alpha^2 + (2 \pi f)^2} \]
Step 1: Recall the PSD is the Fourier transform of the autocorrelation:
\( S_{WN}(f) = \int_{-\infty}^\infty R_{WN}(\tau) e^{-j 2 \pi f \tau} d\tau = \int_{-\infty}^\infty \frac{N_0}{2} \delta(\tau) e^{-j 2 \pi f \tau} d\tau \)
Step 2: Use the sifting property of delta function:
\( S_{WN}(f) = \frac{N_0}{2} e^{-j 2 \pi f \cdot 0} = \frac{N_0}{2} \)
Step 3: Interpretation:
The PSD is constant for all frequencies, indicating that white noise has equal power at every frequency - a flat spectrum. This means white noise contains all frequency components equally.
Answer: \( S_{WN}(f) = \frac{N_0}{2} \), a flat PSD across all frequencies.
Step 1: Recall the output PSD of an LTI system is:
\( S_Y(f) = |H(f)|^2 S_X(f) \)
Step 2: Calculate the magnitude squared of \( H(f) \):
\( |H(f)|^2 = \left| \frac{1}{1 + j 2 \pi f / f_c} \right|^2 = \frac{1}{1 + (2 \pi f / f_c)^2} \)
Step 3: Substitute into output PSD:
\( S_Y(f) = \frac{S_X(f)}{1 + (2 \pi f / f_c)^2} \)
Step 4: Interpretation:
The LTI system acts as a low-pass filter, attenuating higher frequencies and shaping the PSD accordingly.
Answer: The output PSD is
\[ S_Y(f) = \frac{S_X(f)}{1 + (2 \pi f / f_c)^2} \]
Step 1: Check symmetry:
\( R_X(-\tau) = 5 e^{-|- \tau|} = 5 e^{-|\tau|} = R_X(\tau) \). Symmetry holds.
Step 2: Check maximum at zero lag:
\( R_X(0) = 5 e^{0} = 5 \), which is the maximum since \( e^{-|\tau|} \leq 1 \) for all \( \tau \).
Answer: The autocorrelation function is symmetric and attains its maximum value at zero lag, confirming the properties.
When to use: Quickly estimate signal energy from autocorrelation values.
When to use: Simplify autocorrelation computations and check for errors.
When to use: When PSD or autocorrelation is easier to analyze or given in problems.
When to use: Quickly solve noise-related questions without lengthy calculations.
When to use: To avoid complicated convolutions and save time in filtering problems.
Progress tracking is paywalled — subscribe to mark subtopics as understood and save your streak.
Go to practice →
Your Score
Your Rank
| Rank | Name | Score |
|---|
