5.2. Discrete Fourier Transform#

The frequency-domain sampling result given in (5.1) motivates the following definition of \(M\)-point Discrete Fourier Transform (DFT) pair between a signal of length at most \(M\) and the \(M\) frequency domain samples \(X_k = X(e^{j\frac{2\pi k}{M}})\) for \(k=0,1,2,\ldots,M-1\):

Forward DFT:

(5.2)#\[\begin{align} X_k &= \sum_{n=0}^{M-1} x[n] e^{-j\frac{2\pi kn}{M}} & k=0,1,2, \ldots, M-1 \end{align}\]

Inverse DFT:

(5.3)#\[\begin{align} x[n] &= \frac{1}{M} \sum_{k=0}^{M-1} X_k e^{j\frac{2\pi kn}{M}} & n=0,1,2, \dots, M-1 \end{align}\]

Notation

We use the notation \(x[n] \stackrel{M\text{-DFT}}{\longleftrightarrow} X_k\) to denote the mapping between a discrete-time signal of at most length \(M\) and its set of \(M\)-point DFT coefficients \(\{X_k\}_{k=0}^{M-1}\).

5.2.1. DFT properties#

  • Since the DFT coefficients \(X_k\)’s are the frequency-domain samples of the DTFT \(X(e^{j\hat\omega})\), the properties of DTFT (see Section 2.2.2) carry over to DFT with the caveat that we have to consider the periodic extension \(x_M[n]\) instead of \(x[n]\) when the operation involved makes the resulting signal violate the condition that its value must be zero except for \(n=0,1,2,\ldots,M-1\). We consider two examples below.

  • Circular Time Shifting:

    Let \(x[n] \stackrel{\text{DTFT}}{\longleftrightarrow} X(e^{j\hat\omega})\). Then from the time shifting property of DTFT, we have \(x[n-n_0] \stackrel{\text{DTFT}}{\longleftrightarrow} e^{-j\hat\omega n_0} X(e^{j\hat\omega})\). From (5.1), the frequency-domain samples \(\{ e^{-j\frac{2\pi k n_0}{M}} X(e^{j\frac{2\pi k }{M}}) \}_{k=0}^{M-1}\) give the periodic extended signal \(x_M[n-n_0]\). As the length of \(x[n]\) is at most \(M\), the period of \(x_M[n-n_0]\) over \(n=0,1,2,\dots, M-1\) is simply a circularly shifted version of \(x[n]\) by \(n_0\).

    Notation

    We denote the circular shifting of the time index \(n\) by \(n_0\) with the notation \((n-n_0)\bmod M\) or \((n-n_0)_M\). For example, the circularly shifted version of \(x[n]\) will be \(x[(n-n_0)_M]\).

    Since the frequency-domain samples \(\{ e^{-j\frac{2\pi k n_0}{M}} X(e^{j\frac{2\pi k }{M}}) \}_{k=0}^{M-1}\) are by definition the DFT coefficients of the circularly shifted \(x[(n-n_0)_M]\), we have just established the DFT mapping:

    \[\begin{equation*} x[(n-n_0)_M] ~~\stackrel{M\text{-DFT}}{\longleftrightarrow} ~~e^{-j\frac{2\pi k n_0}{M}} X_k \end{equation*}\]

  • Circular convolution:

    Let \(x[n] \stackrel{\text{DTFT}}{\longleftrightarrow} X(e^{j\hat\omega})\) and \(y[n] \stackrel{\text{DTFT}}{\longleftrightarrow} Y(e^{j\hat\omega})\). The convolution property of DTFT gives \(x[n]*y[n] \stackrel{\text{DTFT}}{\longleftrightarrow} X(e^{j\hat\omega}) Y(e^{j\hat\omega})\). Then from (5.1), the frequency-domain samples \(\{ X(e^{j\frac{2\pi k}{M}}) Y(e^{j\frac{2\pi k}{M}})\}_{k=0}^{M-1}\) give the periodic extended signal \((x*y)_M[n]\).

    Let us examine the signal \((x*y)_M[n]\) in detail for the case that the lengths of both \(x[n]\) and \(y[n]\) are at most \(M\):

    (5.4)#\[\begin{split}\begin{align} (x*y)_M[n] &= \sum_{m=-\infty}^{\infty} (x*y)[n+mM] \\ &= \sum_{m=-\infty}^{\infty} \sum_{k=-\infty}^{\infty} x[k]y[n+mM-k] \\ &= \sum_{k=-\infty}^{\infty} x[k] y_M[n-k] \\ &= \sum_{k=0}^{M-1} x[k] y[(n-k)_M] \end{align}\end{split}\]

    The last expression in (5.4) is similar to a convolution sum between \(x[n]\) and \(y[n]\), except that we use the circularly shifted version \(y[(n-k)_M]\) instead. This observation motivates the following definition of circular convolution:

    Notation

    \(\displaystyle x[n] \circledast_M y[n] = \sum_{k=0}^{M-1} x[k] y[(n-k)_M]\) denotes the circular convolution between two signals \(x[n]\) and \(y[n]\) of length at most \(M\).

    With this definition, we get the following DFT mapping:

    \[\begin{equation*} x[n] \circledast_M y[n] ~~\stackrel{M\text{-DFT}}{\longleftrightarrow} ~~ X_k Y_k \end{equation*}\]

    It is easy to check circular convolution has the following properties similar to standard convolution:

    • Commutativity: \(x[n] \circledast_M y[n] = y[n] \circledast_M x[n]\)

    • Associativity: \((x[n] \circledast_M y[n]) \circledast_M z[n] = x[n] \circledast_M (y[n] \circledast_M z[n])\)

    • Distributivity: \(x[n] \circledast_M (\alpha y[n] + \beta z[n] = \alpha x[n] \circledast_M y[n] + \beta x[n] \circledast_M z[n]\)

    • Identity: \(x[n] \circledast_M \delta[n-k] = x[(n-k)_M]\) for \(k=0,1,\ldots, M-1\)

    • Reduction to convolution: Let \(N\) and \(L\) be the lengths of \(x[n]\) and \(y[n]\), respectively. If \(M \geq N+L-1\), then \(x[n] \circledast_M y[n] = x[n]*y[n]\).

5.2.2. DFT property table#

For each property listed below, assume both \(x[n]\) and \(y[n]\) are signals of at most length \(M\), as well as \(x[n] \stackrel{M\text{-DFT}}{\longleftrightarrow} X_k\) and \(y[n] \stackrel{M\text{-DFT}}{\longleftrightarrow} Y(e^{j\hat\omega})\):

Property

Time domain

Frequency domain

Linearity

\(\alpha x[n] + \beta y[n]\)

\(\alpha X_k + \beta Y_k\)

Circular Time Shifting

\(x[(n-n_0)_M]\)

\(e^{-j \frac{2\pi k n_0}{M}} X_k\)

Circular Frequency Shifting

\(x[n] e^{j \frac{2\pi ln}{M}}\)

\(X_{(k-l)_M}\)

Conjugation

\(x^*[n]\)

\(X^*_{M-k}\)

Time Reversal

\(x[M-n]\)

\(X_{M-k}\)

Circular Convolution

\(x[n] \circledast_M y[n]\)

\(X_k Y_k\)

Multiplication

\(x[n] y[n]\)

\(\frac{1}{M} X_k \circledast_M Y_k\)

Parseval Theorem

\(\displaystyle \sum_{n=0}^{M-1} x[n] y^*[n]\)

\(\displaystyle \frac{1}{M} \sum_{k=0}^{M-1} X_k Y^*_k\)

5.2.3. Frequency-domain filtering using DFT#

  • The last property of circular convolution given above allows us to perform convolution \(x[n]*y[n]\) in the frequency domain using DFT between signals of finite lengths:

    1. Let \(N\) and \(L\) be the lengths of \(x[n]\) and \(y[n]\). Choose \(M \geq N+L-1\) (length of \(x[n]*y[n]\)).

    2. Calculate the \(M\)-point DFT of \(x[n]\) to get \(\{X_k\}_{k=0}^{M-1}\).

    3. Calculate the \(M\)-point DFT of \(y[n]\) to get \(\{Y_k\}_{k=0}^{M-1}\).

    4. Calculate the \(M\)-point IDFT of \(\{X_k Y_k\}_{k=0}^{M-1}\) to obtain \(x[n] \circledast_M y[n] = x[n]*y[n]\).

  • Note that direct implementation of the step above, using the DFT and IDFT formulas (5.2) and (5.3), is actually higher than doing convolution directly. Luckily, the FFT algorithms can significantly speed up the calculations of DFT and IDFT; thus making the frequency-domain analysis above much more computationally efficient.