Reconstruction & DAC

Closing the Loop: From Numbers Back to Sound

The previous three lessons traced the journey from continuous signal to discrete numbers: sampling converts time to a sequence of instants, and quantization maps amplitude to a finite set of levels. But a digital audio system must ultimately produce sound — a pressure wave in air. A display must produce light. A radio transmitter must produce a continuous electromagnetic waveform. Somewhere, the discrete sequence of numbers must be converted back into a smooth, continuous analog signal. That reverse process is reconstruction, and the hardware that performs it is the Digital-to-Analog Converter (DAC).

Reconstruction is not merely the inverse of sampling. The original continuous signal existed before sampling; no hardware can restore information that was lost in that process. What reconstruction can do — and what ideal reconstruction achieves — is produce the unique bandlimited signal consistent with the given sample values. When the Nyquist criterion was satisfied during sampling, that unique signal is exactly the original. Reconstruction is therefore deeply tied to the sampling theorem: the theorem not only guarantees that sufficient samples uniquely determine a bandlimited signal, but also provides the mathematical formula for recovering it.

Ideal Reconstruction: Sinc Interpolation

The Whittaker–Shannon interpolation formula states that a bandlimited signal x(t) can be recovered exactly from its samples x[n] = x(nT_s) via:

The sinc function — sinc(t/T_s) = sin(πt/T_s) / (πt/T_s) — is the time-domain impulse response of an ideal low-pass filter with cutoff frequency f_s/2. In the frequency domain, ideal reconstruction is simply low-pass filtering: pass everything below f_s/2, suppress everything above. The filter removes the spectral replicas (images) that appear at multiples of the sampling rate, leaving only the baseband copy of the original spectrum.

The connection is elegant but impractical. An ideal sinc has infinite duration in time and requires non-causal processing — it needs to “see” the future. Real reconstruction must therefore approximate the ideal using causal, finite-length filters. The quality of the approximation is one of the central engineering challenges of DAC design.

Imaging: The Mirror Problem

To understand why a reconstruction filter is needed, consider what a DAC actually outputs before any filtering. The digital sequence x[n] is fed to a circuit that updates its output voltage at each sample period T_s. The output is a staircase waveform: a series of rectangular pulses, each held constant for one sample period, at the amplitude of the corresponding sample. This operation is called a zero-order hold (ZOH).

The ZOH output is continuous in time — no longer a sequence of impulses — but its spectrum contains unwanted spectral copies centered at every multiple of the sampling frequency: f_s, 2f_s, 3f_s, and so on. These copies are called images. If not removed, they appear as distortion in the analog output: for audio, they would be heard as high-frequency artifacts; for communications, they would interfere with adjacent channels.

The ZOH is the crudest possible reconstruction — a flat step held for each sample. One step up is the first-order hold (FOH), which connects consecutive samples with straight line segments instead of flat plateaus, producing a piecewise-linear output. Linear interpolation is the convolution of the sample train with a triangular pulse of width 2T_s — itself the convolution of two rectangular ZOH pulses — so the FOH frequency response is exactly the square of the ZOH's. The cost is one extra sample of delay, because a straight line to the next sample cannot be drawn until that sample has arrived.

The Reconstruction Filter

The reconstruction filter (also called an anti-imaging filter or post-DAC filter) is a low-pass filter applied to the ZOH output. Its job is to:

1. Remove spectral images. All energy above f_s/2 must be suppressed before the signal leaves the digital system. The filter stopband must provide enough attenuation that no image is audible or measurable in the final output. For audio, this typically means −80 dB or better suppression.

2. Compensate for the ZOH roll-off. The staircase approximation attenuates high frequencies within the baseband. A good reconstruction filter includes sinc correction: a boost at higher baseband frequencies that exactly inverts the ZOH attenuation, restoring a flat frequency response across the full audio band.

3. Introduce minimal distortion in the passband. Just like the anti-aliasing filter before the ADC, the reconstruction filter must have flat magnitude and linear phase in the passband to avoid audible coloration or dispersion.

Oversampling: Trading Sample Rate for Filter Complexity

Building an analog reconstruction filter with a steep roll-off is hard. An ideal brick-wall response at exactly f_s/2 requires infinitely many filter poles. A real analog filter achieves only a gradual roll-off, leaving a transition band between the passband edge and the first image. Images in this transition band are not fully suppressed.

The solution, universal in modern DACs, is oversampling: run the DAC at a much higher rate than the nominal sampling frequency, inserting interpolated samples to fill the gaps. If the DAC runs at 8× the audio sample rate, the nearest image is now at 8f_s − f_max instead of f_s − f_max. The images are pushed far up in frequency, and the analog reconstruction filter can be a gentle, low-order design that easily suppresses them.

The interpolation (upsampling) required for oversampling is performed digitally before the DAC. The digital interpolation filter has a sharp, precise cutoff that is impossible to achieve in analog. This is the key insight: shift the hard filtering problem from the analog domain (expensive, imprecise) to the digital domain (cheap, exact). The analog filter that follows needs only a gentle first- or second-order response.

Sigma-Delta DACs: 1-Bit Precision at Extreme Rates

The extreme end of oversampling is the sigma-delta DAC, used in virtually all modern high-fidelity audio systems. Instead of producing a multi-bit output at the Nyquist rate, a sigma-delta DAC uses a 1-bit (or few-bit) output but runs at hundreds of times the audio sampling rate. Each output sample is simply “high” or “low,” but the density of “high” pulses encodes the desired amplitude: a loud signal has more high pulses; a quiet signal has fewer.

The digital modulator that converts multi-bit samples to this 1-bit pulse stream is called a delta-sigma modulator. It uses a feedback loop to push the quantization noise (from reducing many bits to just one) up to high frequencies where it can be removed by the output filter. This technique — called noise shaping — achieves remarkably high signal-to-noise ratios from very coarse quantization: a well-designed sigma-delta DAC can reach 120 dB or more of dynamic range using just 1-bit conversion at a few MHz.

The Complete ADC–DAC Chain

Now all four steps of digital signal processing can be seen as a complete pipeline. The analog input is first anti-alias filtered, then sampled and quantized by the ADC. The digital processor manipulates the samples. The DAC reconstructs a continuous signal, and the reconstruction filter removes images. Four operations, two in analog and two conceptually in digital, together form the backbone of every audio codec, wireless transceiver, and digital instrument in the world.

Each stage has its corresponding artifact when imperfect:

Anti-aliasing filter → passband distortion if the filter is not flat and linear-phase in the signal band.

Sampling (ADC) → aliasing if the Nyquist criterion is violated.

Quantization (ADC) → quantization noise, clipping if bit depth is insufficient or input range is exceeded.

DAC + reconstruction filter → imaging, roll-off, jitter if the filter is inadequate, the oversampling ratio is low, or the sample clock is unstable.

Jitter: When the Clock Wobbles

One reconstruction artifact deserves special mention: jitter. Ideal reconstruction assumes samples are produced at perfectly uniform intervals T_s. In practice, the DAC clock has small random timing errors — samples arrive slightly early or late. This timing uncertainty causes the output waveform to be distorted even if the amplitude values are perfectly correct.

The effect of jitter is equivalent to amplitude noise: if the DAC output is a sinusoid at frequency f₀ and the clock has rms timing jitter σ_t, the resulting SNR degradation is approximately:

Jitter is minimized by using low-noise crystal oscillators, phase-locked loops (PLLs) with tight loop bandwidths, and careful board layout to isolate the clock from digital noise sources. In professional audio, a separate word-clock signal distributes timing across multiple devices to keep all converters phase-locked.