DFT and IDFT Equations: Derivation & Analysis | DSP Lecture 4

Circular convolution differs from linear convolution because it wraps around the time-domain samples due to the periodic nature of the Discrete Fourier Transform (DFT), meaning the multiplication of two DFTs equals their circular convolution; to obtain valid linear convolution results, one must discard the wrapped-around samples by zero-padding the input signals to at least the length of their linear convolution output before performing the circular convolution.

Circular convolution is a fundamental signal processing operation where two sequences are convolved with periodic boundary conditions. Unlike linear convolution which produces an output of length L+M-1, circular convolution produces an output of the same length as the input sequences. The operation uses the Discrete Fourier Transform (DFT) where the frequency variable omega is sampled at intervals of 2π/N. The mathematical formula is: y(n) = Σ(k=0 to N-1) [x1(k) × x2((n-k) mod N)], where n varies from 0 to N-1. The modulo operation ensures indices wrap around when they exceed N-1 or become negative. This periodic nature causes sequences to overlap and wrap around, which is the key characteristic distinguishing circular convolution from linear convolution.

This segment covers the difference between linear and circular convolution, which are used for different types of signals. Linear convolution is used for aperiodic signals and produces an output length of L1 + L2 - 1. Circular convolution is used for periodic signals and produces an output length equal to the maximum of L1 and L2. The instructor explains that the key difference is that circular convolution wraps around the output, while linear convolution extends the output. When working with periodic signals, circular convolution is the appropriate operation. The instructor demonstrates how to calculate both types of convolution using examples, showing the step-by-step process of applying these operations to discrete-time signals.

In DFT-based processing, circular convolution is used instead of linear convolution. Circular convolution of two sequences x1[n] and x2[n] of length N is defined as y[n] = Σ(m=0 to N-1) x1[m] * x2[(n-m) mod N]. This operation is equivalent to multiplying their DFTs: Y[k] = X1[k] * X2[k]. The circular nature means the result wraps around with period N. This is different from linear convolution, which produces a sequence of length N1 + N2 - 1. When N equals N1 + N2 - 1, circular and linear convolution produce identical results.

Circular convolution is a fundamental operation in digital signal processing that differs from linear convolution. It operates on periodic signals in the DFT domain, where signals repeat every 2π radians. The definition involves summing products of sequence elements with modulo indexing: x1(n) ⊛ x2(n) = Σ (m=0 to N-1) x1(m) × x2((n-m) mod N). The key DFT convolution property states that circular convolution in the time domain corresponds to multiplication in the frequency domain: DFT{x1(n) ⊛ x2(n)} = X1(k) × X2(k). This property simplifies complex convolution operations into efficient multiplication operations, making it essential for signal processing applications.

The Fast Fourier Transform (FFT) is an efficient algorithm that reduces the computational complexity of calculating the Discrete Fourier Transform (DFT) from O(N²) to O(N log₂N) by exploiting the special properties of complex exponentials and recursively decomposing the DFT into smaller sub-problems through decimation-in-time, where the input sequence is split into even and odd indexed subsequences, each processed separately and then combined with appropriate twiddle factors (WN^k).

This tutorial demonstrates how to compute an 8-point Discrete Fourier Transform (DFT) using the Radix-2 Decimation-in-Time (DIT) Fast Fourier Transform (FFT) algorithm. The process involves four key steps: (1) arranging the input sequence in bit-reversed order, (2) performing butterfly computations at three stages with appropriate twiddle factors (W_N^k), and (3) calculating the final frequency-domain output. For the given input sequence x(n) = [1, 2, 4, 8, 16, 32, 64, 128], the resulting DFT output is X(k) = [255, 48.63+j166.05, -51+j102, -78.63+j46.05, -85, -78.63-j46.05, -51-j102, 48.63-j166.05]. The FFT algorithm significantly reduces computational complexity compared to direct DFT calculation by exploiting symmetries in the transform.

The Radix-2 Decimation-in-Time Fast Fourier Transform (FFT) algorithm efficiently computes the Discrete Fourier Transform by recursively breaking down a sequence of length N into two smaller sequences of length N/2, using a butterfly structure that combines results from smaller transforms to achieve O(N log N) computational complexity instead of O(N²) for direct DFT computation.

The FFT algorithm uses a butterfly structure where DFT samples are split into pairs with shared W_N components. A two-sample DFT requires only 2 MAC operations. A four-sample DFT combines two two-sample DFTs using a butterfly with W_4^0 and W_4^1 multipliers, requiring 8 MAC operations. An eight-sample DFT recursively combines four two-sample DFTs into four-sample DFTs, then into one eight-sample DFT, requiring 24 MAC operations. The general formula is N × log₂(N) MAC operations. For a 1024-sample signal, this reduces computation from over 1 million to only 10,000 MAC operations—a 100-fold improvement.

The Radix-2 FFT algorithm reduces DFT computational complexity from O(N²) to O(N log N) by recursively decomposing the DFT into smaller DFTs of length N/2, leveraging properties of the twiddle factor W_N and using an in-place computation with bit-reversed ordering, where each stage combines pairs of inputs through butterfly operations that perform additions, subtractions, and multiplications by complex factors.

This comprehensive lesson covers the fundamental problem of spectral leakage in DFT analysis and its solution through windowing. The DFT assumes signals contain whole numbers of periods over the analysis interval; when violated, artificial discontinuities require many high-frequency components, spreading energy across bins. Window functions solve this by multiplying signals to dampen edges, with the sine window w(n)=sin(πn/N) and Hanning window w(n)=0.5×(1-cos(2πn/N)) providing progressively better leakage reduction. Practical application to STFT demonstrates how windowing concentrates energy into correct frequency bins, enabling accurate spectral analysis of real-world signals with arbitrary frequencies.

Applying a window function (such as Hamming) before performing the Discrete Fourier Transform (DFT) reduces spectral leakage by modifying the implicit rectangular window; this works because multiplying a signal by a window in the time domain corresponds to convolving its spectrum with the window's frequency response, thereby spreading the signal's energy into a wider main lobe but reducing the amplitude of side lobes that cause spectral leakage.

This section explains the phenomenon of spectral leakage and how window functions mitigate its effects in FFT analysis. When applying DFT to real-world signals, the signal is treated as periodic, but real signals don't naturally repeat, causing phase discontinuities at window boundaries. These discontinuities generate spurious spectral components called leakage. Window functions multiply the signal data before transformation, shaping the signal segment to reduce leakage. Different window types offer varying suppression levels: rectangular window provides minimal suppression (~-16 dB) but best resolution; Hamming, Hanning, and Blackman-Harris windows provide progressively better suppression (up to ~-90 dB); Kaiser window offers adjustable characteristics. The trade-off involves balancing frequency resolution against leakage reduction, with the choice depending on specific analysis requirements.

Spectral leakage occurs when the input signal does not contain an integer number of cycles within the analysis window (N samples). In such cases, the DFT cannot find an exact match with any basis function, causing energy to spread across multiple adjacent frequency bins rather than appearing as a single sharp peak. For example, a signal with 2.3 cycles over 8 samples will show significant values at bins near k=2 and k=3, with progressively smaller values at other bins. This phenomenon, called spectral spreading, makes it difficult to precisely identify sinusoidal components and is a fundamental limitation of the DFT when analyzing signals with non-integer cycle counts.

Spectrum leakage occurs because the DFT treats all signals as periodic, causing discontinuities at segment boundaries when real-world signals don't align with their periods. These discontinuities cause energy to leak into adjacent frequency bins, producing side lobes instead of clean impulses. Windowing techniques (like the Hamming window) multiply signal segments by tapering functions that gradually reduce amplitude to zero at both ends, eliminating sharp discontinuities. This reduces leakage at the cost of slightly broadening frequency responses and reducing signal amplitude. The lecture emphasizes that windowing is essential for accurate spectral analysis of non-periodic signals in practice, representing a fundamental trade-off between signal modification and measurement accuracy.

When processing long signals through filters, dividing into frames of length l produces outputs of length l + m - 1, causing length mismatch. Two solutions exist: (1) Overlap-Save: Pad each frame with m-1 zeros at the beginning, process, discard the first m-1 output samples, and keep the remaining l samples. Consecutive frames overlap by m-1 samples. (2) Overlap-Add: Pad each frame with m-1 zeros at the end, process, and add overlapping output portions. Both methods ensure the final output length matches the input length.

The overlap save and overlap add methods are two techniques used in digital signal processing to efficiently filter long data sequences on devices with limited memory. Both methods divide the input sequence into smaller blocks, apply DFT-based filtering to each block, and then combine the results. In the overlap save method, each block contains M-1 samples from the previous block plus L new samples, and the first M-1 samples of each output block are discarded before concatenation. In the overlap add method, each block contains only L new samples with zeros appended to reach the required length, and the blocks are overlapped and added together with the zeros. These methods enable efficient filtering of long sequences by reducing memory requirements while maintaining computational accuracy.

The overlap save method is a technique for filtering long data sequences using DFT by dividing the input into overlapping blocks of length n = l + m - 1 (where l is new data length and m is filter length), zero-padding the impulse response, multiplying DFTs, discarding the first m-1 corrupted points due to aliasing, and fitting the remaining l points together to form the overall output sequence.

Real-time audio processing requires block-based convolution methods. The overlap-add scheme processes input in blocks, performs fast convolution with the filter, stores results, and adds successive output parts to subsequent blocks. Transform length must be at least input block length plus filter length minus one. The overlap-save scheme stores input blocks instead of results, computes circular convolution, and outputs only the last b valid samples. Both methods handle circular convolution artifacts through strategic zero-padding and block management, enabling continuous real-time filtering without storing all previous results.

The Overlap-Add method is a technique for filtering long duration sequences by dividing the input signal into overlapping segments, performing circular convolution with the filter, and adding the last M-1 points of each segment to the beginning of the next segment to produce the final output sequence.