## Interpolation and energy amplification

Consider the following interpolator [1]

where $x_u = (\uparrow L) (x)$ is the upsampling of sequence $x$ by an integer factor of $L$, and where $H(z)$ is the transfer function of an ideal discrete-time LTI lowpass filter with gain $L$ and cutoff frequency $\pi/L$. What is the relation between the energies of signals $x$ and $y$?

__________

My solution:

It can be easily shown [1] that the interpolator’s input-output relationship is $Y (z) = H(z) X (z^L)$. Hence, evaluating the Z-transforms on the unit circle, we obtain

$Y (e^{j \omega}) = H(e^{j \omega}) X (e^{j \omega L})$.

The energy of the interpolator’s output signal is

$\mathcal{E}_y := \displaystyle \sum_{n \in \mathbb{Z}} | y(n) |^2 = \frac{1}{2 \pi} \int_{-\pi}^{\pi} |Y (e^{j \omega})|^2 d \omega$

where we used Parseval’s theorem for discrete-time signals. Thus,

$\mathcal{E}_y = \displaystyle\frac{1}{2 \pi} \int_{-\pi}^{\pi} |H(e^{j \omega})|^2 |X (e^{j \omega L})|^2 d \omega$.

$H(z)$ is an ideal lowpass filter with the following frequency response

$H(e^{j \omega}) = \begin{cases} L, & |\omega| \leq \pi/L\\ 0, & \pi/ L< |\omega| \leq \pi\end{cases}$

and the following sinc impulse response

$h(n) = \displaystyle\frac{1}{2 \pi} \int_{-\pi}^{\pi} H(e^{j \omega}) e^{j \omega n} d \omega = \displaystyle \frac{\sin(\frac{n \pi}{L})}{\frac{n \pi}{L}} =: \text{sinc} \left(\frac{n \pi}{L}\right)$.

Please do note that this is a non-causal filter with an impulse response of infinite length. We can then write the interpolator’s output signal’s energy as follows

$\mathcal{E}_y = \displaystyle\frac{1}{2 \pi} \int_{-\pi}^{\pi} |H(e^{j \omega})|^2 |X (e^{j \omega L})|^2 d \omega = \displaystyle\frac{L^2}{2 \pi} \int_{-\pi/L}^{\pi/L} |X (e^{j \omega L})|^2 d \omega$.

Performing a change of variable, $\theta = L \omega$, we obtain

$\mathcal{E}_y = \displaystyle\frac{L^2}{2 \pi} \int_{-\pi/L}^{\pi/L} |X (e^{j \omega L})|^2 d \omega = \displaystyle\frac{L}{2 \pi} \int_{-\pi}^{\pi} |X (e^{j \theta})|^2 d \theta = L \mathcal{E}_x$

where

$\mathcal{E}_x := \displaystyle \sum_{n \in \mathbb{Z}} | x(n) |^2 = \frac{1}{2 \pi} \int_{-\pi}^{\pi} |X (e^{j \omega})|^2 d \omega$

is the energy of the input signal. Since $\mathcal{E}_y = L \mathcal{E}_x$, we can conclude that the interpolator amplifies the input signal’s energy by a factor of $L$, which is somewhat intuitive.

We considered the case where signals are bi-infinite sequences and $H(z)$ is a non-causal ideal lowpass filter. It would be interesting to consider the more realistic case where the signals are finite sequences and $H(z)$ is a causal FIR filter.

__________

References

[1] Alan V. Oppenheim, Ronald W. Schafer, Discrete-Time Signal Processing, 2nd edition, Prentice-Hall, 1999.

### 2 Responses to “Interpolation and energy amplification”

1. eryksun Says:

It’s inaccurate to call this energy amplification. When one calculates discrete-time energy, it’s in WattÂ·samples. The result only truly becomes a physical unit of energy after scaling by the sample time, i.e.

$\text{Watt} \times \text{samples} \times \displaystyle\frac{\text{seconds}}{\text{sample}} = \text{Watt} \times \text{seconds} = \text{Joules}$.

While there are $L$ times as many WattÂ·samples, the sample time is also reduced by $L$. The gain of $L$ on the low pass filter is there to compensate for the loss due to the expander (i.e. inserting zeros). It’s not amplification.

• Rod Carvalho Says:

As an EE, I understand your objection. Nonetheless I have to disagree. It is energy amplification because I defined energy as

$\mathcal{E}_x := \displaystyle \sum_{n \in \mathbb{Z}} | x(n) |^2$.

Nowhere did I claim that this energy was physical. Interpolation is part of Numerical Analysis, too. It’s wider than DSP. And not all sequences are obtained by sampling continuous-time signals.

Having said that, your comment got me interested in the relation between the physical energies of the interpolator’s input and output.