Reconstruction of Signals

Introduction

Reconstruction is the process of converting discrete samples back into a continuous-time signal — essentially the reverse of sampling. Every time a digital audio file plays through a speaker, every time a digital control system drives a motor, and every time a computer generates an analog output, signal reconstruction is happening. The Sampling Theorem guarantees that if the signal was properly sampled (above the Nyquist rate), perfect reconstruction is theoretically possible. However, the practical methods used to achieve reconstruction involve important engineering trade-offs.

The ideal reconstruction uses sinc interpolation, which produces mathematically perfect results but is physically unrealizable because it requires knowledge of samples from both the infinite past and infinite future. Practical systems use causal approximations — zero-order hold, linear interpolation, or higher-order methods — followed by analog reconstruction filters that smooth out the resulting staircase or piecewise-linear waveforms.

Ideal Reconstruction: Shannon Interpolation

The Shannon-Whittaker interpolation formula provides the mathematically exact reconstruction:

$$x(t) = \sum_{n=-\infty}^{\infty} x[n] \cdot \text{sinc}\left(\frac{t - nT_s}{T_s}\right)$$

where $\text{sinc}(u) = \frac{\sin(\pi u)}{\pi u}$ and $T_s$ is the sampling period.

Each sample $x[n]$ generates a sinc pulse centered at $t = nT_s$. The continuous signal is the superposition of all these sinc pulses. The sinc function is special because:

• It equals 1 at its center ($t = nT_s$) and 0 at all other sample points
• Its Fourier Transform is a perfect rectangular function (ideal low-pass)
• It has infinite temporal extent (extending to $\pm\infty$)

Frequency-Domain Interpretation

Ideal reconstruction is equivalent to passing the sampled signal through an ideal low-pass filter with:

• Cutoff frequency: $f_c = f_s/2$
• Gain in passband: $T_s$
• Perfectly flat passband and zero stopband

$$H_{ideal}(j\omega) = \begin{cases} T_s, & |\omega| < \omega_s/2 \\ 0, & |\omega| > \omega_s/2 \end{cases}$$

This filter selects the baseband spectral replica and removes all the higher-frequency copies created by sampling.

Why Ideal Reconstruction is Impractical

The sinc function creates three fundamental problems:

1. Non-causal: The sinc extends to $-\infty$, meaning reconstruction at time $t$ requires future samples (not yet available in real-time systems)
2. Infinite duration: Perfect reconstruction requires summing contributions from infinitely many samples
3. Slow decay: The sinc decays as $1/t$, meaning truncation introduces significant error

These issues motivate practical approximations.

Practical Method 1: Zero-Order Hold (ZOH)

The zero-order hold is the simplest practical reconstruction: hold each sample value constant until the next sample arrives.

$$x_{ZOH}(t) = x[n], \quad nT_s \leq t < (n+1)T_s$$

The output is a staircase approximation. In the frequency domain:

$$H_{ZOH}(j\omega) = T_s \cdot \text{sinc}\left(\frac{\omega T_s}{2\pi}\right) \cdot e^{-j\omega T_s/2}$$

The ZOH introduces:

• Amplitude distortion: The sinc envelope attenuates higher frequencies (droop)
• Half-sample delay: The $e^{-j\omega T_s/2}$ term introduces a delay of $T_s/2$

At $f = f_s/2$ (the Nyquist frequency), the sinc attenuation is $2/\pi \approx 0.637$ (about -3.9 dB). This droop must be compensated by either a digital pre-emphasis filter or an analog post-filter.

Practical Method 2: First-Order Hold (Linear Interpolation)

First-order hold connects adjacent samples with straight lines:

$$x_{FOH}(t) = x[n] + \frac{x[n+1] - x[n]}{T_s}(t - nT_s), \quad nT_s \leq t < (n+1)T_s$$

This produces a piecewise-linear waveform that is smoother than ZOH. The frequency response is:

$$H_{FOH}(j\omega) = T_s \cdot \text{sinc}^2\left(\frac{\omega T_s}{2\pi}\right)$$

The squared sinc provides better attenuation of spectral images (faster rolloff) but more droop in the passband.

Practical Method 3: Higher-Order Interpolation

Cubic spline interpolation: Fits cubic polynomials between samples with continuous first and second derivatives. Produces very smooth results with modest computational cost.

Windowed sinc: Truncates the ideal sinc to a finite window (e.g., 64 or 128 taps) and applies a tapering window function. This is the approach used in high-quality audio resampling.

Polynomial (Lagrange) interpolation: Uses $N$ surrounding samples to fit a polynomial of degree $N-1$. More samples give better approximation of the ideal sinc.

The Reconstruction Filter

After the ZOH or FOH generates a continuous (but imperfect) waveform, an analog reconstruction filter removes the spectral images and compensates for droop:

$$x_{reconstructed}(t) = x_{hold}(t) * h_{filter}(t)$$

Requirements for the reconstruction filter:

• Pass the baseband signal (0 to $f_{max}$) with minimal distortion
• Attenuate spectral images centered at $f_s, 2f_s, 3f_s, \ldots$
• Compensate for the sinc-shaped droop introduced by the hold circuit

In practice, this is implemented as a Butterworth, Chebyshev, or elliptic low-pass filter with cutoff near $f_s/2$.

Oversampling in DACs

Modern digital-to-analog converters use oversampling to simplify the reconstruction filter:

1. Digital upsampling: Insert zeros between original samples and apply a digital low-pass filter to interpolate (e.g., 4× or 8× oversampling)
2. High-rate DAC: Convert at the oversampled rate
3. Simple analog filter: Because the spectral images are now pushed far away (at $4f_s$ or $8f_s$), a gentle analog filter suffices

This technique is used in CD players, audio DACs, and data communication systems. The digital interpolation filter does the precision work; the analog filter handles only coarse smoothing.

Worked Example

Problem: Samples $x[n] = \{1, 2, 3, 2, 1\}$ at $T_s = 1$ ms need to be reconstructed. Describe the ZOH and ideal outputs.

ZOH output: Staircase holding each value for 1 ms. Output is 1V from 0-1ms, 2V from 1-2ms, 3V from 2-3ms, 2V from 3-4ms, 1V from 4-5ms.

Ideal output: Smooth curve passing through all five sample points, generated by: $$x(t) = \sum_{n=0}^{4} x[n] \cdot \text{sinc}\left(\frac{t - n \cdot 0.001}{0.001}\right)$$

The ideal output is bandwidth-limited and smooth, with no sharp transitions.

Key Takeaways

• Ideal reconstruction uses sinc interpolation: mathematically perfect but physically unrealizable
• Zero-order hold (staircase) is simplest but introduces sinc-shaped passband droop
• Linear interpolation (first-order hold) is smoother but still requires post-filtering
• An analog reconstruction filter removes spectral images and compensates for hold distortion
• Oversampling DACs push images far from baseband, allowing simple analog filters
• The quality of reconstruction depends on how closely the practical method approximates the ideal sinc